NAVIGATING AI

COMPLIANCE
PART 2: RISK MITIGATION
STRATEGIES FOR SAFEGUARDING
AGAINST FUTURE FAILURES

MARIAMI TKESHELASHVILI
TIFFANY SAADE
MARCH 2025
Navigating AI Compliance, Part 2
Risk Mitigation Strategies for Safeguarding Against Future Failures

March 2025
Author: Mariami Tkeshelashvili and Tiffany Saade
Design: Sophia Mauro

The Institute for Security and Technology and the authors of this report invite free use of
the information within for educational purposes, requiring only that the reproduced material
clearly cite the full source.

Copyright 2025, The Institute for Security and Technology

Printed in the United States of America
About the Institute for Security
and Technology
Uniting technology and policy leaders to create actionable
solutions to emerging security challenges
Technology has the potential to unlock greater knowledge, enhance our collective capabilities,
and create new opportunities for growth and innovation. However, insecure, negligent, or
exploitative technological advancements can threaten global security and stability. Anticipating
these issues and guiding the development of trustworthy technology is essential to preserve
what we all value.

The Institute for Security and Technology (IST), the 501(c)(3) critical action think tank, stands at
the forefront of this imperative, uniting policymakers, technology experts, and industry leaders
to identify and translate discourse into impact. We take collaborative action to advance
national security and global stability through technology built on trust, guiding businesses and
governments with hands-on expertise, in-depth analysis, and a global network.

We work across three analytical pillars: the Future of Digital Security, examining the systemic
security risks of societal dependence on digital technologies; Geopolitics of Technology,
anticipating the positive and negative security effects of emerging, disruptive technologies on
the international balance of power, within states, and between governments and industries;
and Innovation and Catastrophic Risk, providing deep technical and analytical expertise on
technology-derived existential threats to society.

Learn more: https://securityandtechnology.org/

Acknowledgments
This work is inherently collaborative. As researchers, conveners, and facilitators, the Institute
for Security and Technology (IST) is immensely grateful to the members of the AI Risk
Reduction multi stakeholder working group for their insights, dedication, willingness to engage
in honest and healthy debate, and the time that each of them generously volunteered to this
effort.

We are also immensely grateful for the generous support of Patrick J. McGovern Foundation,
whose funding allowed us to continue this project through the lens of IST’s Applied Trust &
Safety program.

AI is too vast a set of tools, capabilities, and communities for any one organization to manage
the risks and opportunities on its own. This effort reflects the cross-sectoral, public-private
efforts needed more broadly across the ecosystem to ensure AI is beneficial for us all. We
extend our gratitude to the following experts who contributed to this paper by providing their
feedback, guidance, and participation in the multi stakeholder meetings:

» Chloe Autio » Brian Judge

» Henriette Cramer » Monica Lopez
» Matthew da Mota » Alexander Reese
» Hadassah Drukarch » Alyssa Lefaivre Škopac
» Avijit Ghosh » Justin Sherman
» Kamya Jagadish » Peter Slattery
» Katherine Johnson » Akash Wasil
Finally, the authors extend their gratitude to Steve Kelly and Philip Reiner for the support and
strategic guidance they provided in the drafting and refining of this report.
Contents
Executive Summary......................................................................................1
Recap of Navigating AI Compliance, Part 1 ................................................. 3
Introduction................................................................................................ 4
Methodology............................................................................................... 5
The AI Lifecycle Stages....................................................................................................................7
Return on Investment for Implementing Strong Compliance Practices 8
Risk Mitigation Strategies for Safeguarding Against Future Failures 10
Data Collection and Preprocessing ................................................................................................11
Model Training and Evaluation ...................................................................................................... 14
Model Application.......................................................................................................................... 17
User Interaction .............................................................................................................................18
Ongoing Monitoring and Maintenance..........................................................................................19
Conclusion................................................................................................ 20
Executive Summary
Historical trends do not wholly dictate the future of AI. While the first installment of this report
acknowledged the importance of historic lessons, we can make deliberate choices to shape
what comes next. We hope that “Navigating AI Compliance, Part 2: Safeguarding Against
Future Failures” will guide decision-makers in fostering societal trust in AI systems, all while
preventing the repetition of past mistakes.

This report, the second in a two-part series, presents 39 risk mitigation strategies for avoiding
institutional, procedural, and performance failures of AI systems (see Risk Mitigation Strategies
for Safeguarding Against Future Failures). These strategies aim to enhance user trust in AI
systems and maximize product utilization. AI builders and users, including AI labs, enterprises
deploying AI systems, as well as state and local governments, can use and implement a
selection of the 22 technical and 17 policy-oriented risk mitigation strategies presented in this
report according to their needs and risk thresholds.

Through implementing these practices, organizations building and utilizing AI systems not
only reduce regulatory risk exposure and build user trust for their product, but they could also
attract top talent, gain a competitive edge, enhance their financial performance, and increase
the lifetime value of their solutions. Based on our research and the results of stakeholder
engagement, we emphasize to AI builders and users the following nine recommendations
from the complete list of 39:

» Implement proportional compliance measures for high-impact AI applications. AI

builders and users should consider which compliance measures are most appropriate
for their work, especially when building or deploying AI systems in sensitive or high-
impact areas. This consideration should be proportional to factors such as the intended
use, potential risks, and application domain—ranging from entertainment to critical
sectors like national security, healthcare, and finance.

» Acknowledge and address acceptable risks in AI development and deployment.

Unintended consequences are not to be confused with compliance failure. Still, these
unplanned effects should be acknowledged by developers, builders, and regulators as
they consider thresholds of acceptable tolerance for the enhanced risks associated with
exposed attack surfaces and features or functionalities of AI that are not yet thoroughly
understood or anticipated.

March 2025 securityandtechnology.org 1

» Prioritize data management and privacy practices to maintain user trust. Implementing
proper data management and privacy-enhancing practices will protect user rights,
maintain trust, and comply with data protection regulations. Measures such as privacy-
preserving technologies, content provenance features, and user consent mechanisms
can alleviate procedural failures.

» Implement robust cybersecurity controls for AI infrastructure protection and

enhanced reliability. Cybersecurity controls, red-teaming, fail-safe mechanisms, and
other techniques protect AI systems from attacks and strengthen their reliability in
various scenarios. Security guardrails may alleviate or preempt both performance and
procedural failures.

» Utilize safety and risk assessments to proactively mitigate AI harms. Safety and
risk assessment procedures, such as incident reporting frameworks and AI safety
benchmarks at different stages of the lifecycle, identify and mitigate possible harms
before they occur–potentially mitigating both procedural and performance failures.

» Design and implement compliance and AI literacy training for staff. Training should be
mandatory for all staff members involved in the AI supply chain, from data providers to
model developers and deployers. All staff members utilizing AI tools in some manner
should also obtain a minimum set of AI literacy skills through the training.

» Build trust by implementing transparency mechanisms. Transparency and

interpretability mechanisms such as model cards, data cards, and disclosure frameworks
are necessary to build user and stakeholder trust, facilitate accountability, and enable
informed decision-making.

» Enhance AI explainability and disclosure frameworks to improve understanding of

system behavior. Efforts to increase the explainability of AI systems, supplemented
with disclosure frameworks for model evaluation, allow both builders and users to
better understand the behavior patterns and outputs of these systems and potentially
safeguard against performance failures.

» Employ strategies for non-discriminatory AI. Bias mitigation strategies across model
training, data collection, and ongoing monitoring and maintenance, in addition to
adversarial debiasing, can prevent performance failures and help to ensure fairness
while preventing discriminatory outcomes in AI systems.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 2
Recap of Navigating AI
Compliance, Part 1: Tracing
Failure Patterns in History
History often rhymes with and echoes through the present and future. Through this lens, the
first installment of this two-part report series examined past compliance failures across various
industries–from nuclear power to financial services–as a source of definitions, frameworks,
and lessons learned to help AI builders and users navigate today’s complex compliance
landscape.1 Our analysis of eleven case studies from AI-adjacent industries revealed three
distinct categories of failure:
» Institutional failures stem from a lack of executive commitment to create a culture
of compliance, establish necessary policies, or empower success through the
organizational structure, leading to foreseeable failures.
» Procedural failures are the result of a misalignment between an institution’s established
policies and its internal procedures and staff training required to adhere to those
policies.
» Performance failures result when employees fail to follow an established process, or an
automated system fails to perform as intended, leading to an undesirable result.

By studying failures across sectors, we uncovered critical lessons about risk assessment,
safety protocols, and oversight mechanisms that can guide AI innovators in this era of
rapid development. One of the most prominent risks is the tendency to prioritize rapid
innovation and market dominance over safety. The case studies demonstrated a crucial need
for transparency, robust third-party evaluation and verification, and comprehensive data
governance practices, among other safety and security measures.

Though today’s AI regulatory landscape remains fragmented, we identified five main sources
of AI governance—laws and regulations, guidance, norms, standards, and organizational
policies—to provide AI builders and users with a clear direction for the safe, secure, and
responsible development of AI. Therefore, we defined “compliance failure” within the AI

1 Mariami Tkeshelashvili and Tiffany Saade, “Navigating AI Compliance, Part 1: Tracing Failure Patterns in History,” Institute for
Security and Technology, December 2024, https://securityandtechnology.org/wp-content/uploads/2024/12/Navigating-AI-
Compliance.pdf.

March 2025 securityandtechnology.org 3

ecosystem as the failure to align with and adhere to any of these governance mechanisms,
whether publicly announced or confidential.

Part 1 of this report series concluded by addressing AI’s unique compliance issues stemming
from its ongoing evolution and complexity. Ambiguous AI safety definitions and the rapid
pace of development challenge efforts to govern it—including AI’s adoption within regulated
industries—while interpretability challenges hinder the development of compliance
mechanisms. Furthermore, the rapid advent of agentic AI will introduce added complexity and
blur the lines of liability in an increasingly automated world.

Introduction
As illustrated in the first of this two-part report, any technology can fail and cause harm. But
failure of a technology product that has achieved ubiquity in the marketplace can generate
magnified effects—which is the essence of concentration risk. As AI quickly trends toward
ubiquity, the risks of AI system failures and their ripple effects are further magnified by a trend
toward AI autonomy. It is therefore all the more important to manage these risks and alleviate,
pre-empt, and avoid future failures.

How exactly can AI builders and users defend against future failure risks? What are the
benefits of proactively implementing compliance practices? This report aims to:

» Provide AI builders with technical and policy-oriented risk mitigation strategies for
avoiding compliance failures in the future. AI builders are defined in Part 1 of this report
series as “individuals or organizations responsible for developing the models including
AI labs, startups, and tech companies.”2

» Provide AI users with technical and policy-oriented risk mitigation strategies for
responsible deployment of AI systems. AI users are defined in Part 1 as “all other entities
who deploy or utilize the technology, including enterprises integrating AI systems into
their services and internal operations.”3

» Illuminate the various ways in which sound compliance practices can generate return on
investment (ROI).

This report’s proposed risk mitigation strategies are inspired by lessons learned from past
compliance failures noted in Part 1 and co-created by the working group members listed in the

2 Mariami Tkeshelashvili and Tiffany Saade, “Navigating AI Compliance, Part 1.”

3 Mariami Tkeshelashvili and Tiffany Saade, “Navigating AI Compliance, Part 1.”

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 4
acknowledgements section above.4 By developing an actionable compliance pathway for AI
builders and users at each stage of the AI lifecycle, we aim to help bridge the gap between
the drive for AI innovation in global markets and the desire to manage risk.

Methodology
Our research relied on lessons learned from the historical case studies presented in Part 1 of
this report; investigation of databases that reflect the current state of compliance issues within
the AI ecosystem; and over 20 expert interviews with AI labs, tech industry stakeholders,
machine learning engineers, AI governance and policy experts, compliance officers,
attorneys, university-based AI research centers, AI ethicists, and independent researchers.
Complementing this research, IST convened two multi-stakeholder, closed-door discussions
with our AI Risk Reduction working group to gather further insights and agree on the final list
of risk mitigation strategies.

In order to integrate existing AI governance frameworks into our thinking, we analyzed a set of
AI norms, standards, and regulations—both binding and voluntary—to distill the main themes
and patterns for technical and policy mitigation strategies across the AI lifecycle. The sources
we integrated are: voluntary commitments such as the Hiroshima Process;5 the Organization
for Economic Co-operation and Development’s (OECD’s) AI Framework;6 United Kingdom
AI Framework;7 work of the Coalition for Content Provenance and Authenticity (C2PA);8
National Institute of Standards and Technology (NIST) AI Risk Management Framework;9 ISO/
IEC standard 42001;10 and binding regulatory frameworks such as the European Union’s AI
Act11 and General Data Protection Regulation (GDPR).12 Additionally, we integrated relevant

4 Mariami Tkeshelashvili and Tiffany Saade, “Navigating AI Compliance, Part 1.”

5 G7, “Hiroshima Process International Guiding Principles for Organizations Developing Advanced AI Systems,” G7 2023 Hiroshima
Summit, October 30, 2023, https://www.soumu.go.jp/hiroshimaaiprocess/pdf/document04_en.pdf.
6 OECD, “The OECD Artificial Intelligence (AI) Principles,” OECD.AI Policy Observatory, 2019, https://oecd.ai/en/ai-principles.
7 UK Government, “National AI Strategy,” Department for Science, Innovation and Technology, Office for Artificial Intelligence,
Department for Digital, Culture, Media & Sport and Department for Business, Energy & Industrial Strategy, September 22, 2021,
https://www.gov.uk/government/publications/national-ai-strategy.
8 Coalition for Content Provenance and Authenticity, “Guiding Principles - C2PA,” 2024, https://c2pa.org/principles/.
9 National Institute of Standards and Technology, “AI Risk Management Framework,”NIST AI 100-1, January 2023, https://doi.
org/10.6028/nist.ai.100-1.
10 International Standards Organization (ISO), “ISO/IEC DIS 42001,” 2023, https://www.iso.org/standard/81230.html.
11 European Parliament, “EU AI Act: First Regulation on Artificial Intelligence,” June 8, 2023, https://www.europarl.europa.eu/topics/
en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence.
12 “Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with
regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data
Protection Regulation),” Official Journal of the European Union 119/1 (May 4, 2016), https://eur-lex.europa.eu/legal-content/EN/TXT/
PDF/?uri=CELEX:32016R0679.

March 2025 securityandtechnology.org 5

elements from IEEE standards 700013 and 7002,14 and specific ISO standards which are not
exclusive to AI systems but establish important standards for ethical system design and data
privacy.

The risk mitigation strategies presented in this report both leverage and are aligned to IST’s
previously articulated AI Lifecycle Framework, which breaks down the complex process of AI
development into manageable stages.15 This structured approach ensures a comprehensive
understanding of each phase, making it easier to develop and implement specific risk
mitigation strategies.

13 Institute of Electrical and Electronics Engineers (IEEE), “IEEE Standard Model Process for Addressing Ethical Concerns during System
Design,” IEEE 7000-2021, September 15, 2021, https://standards.ieee.org/ieee/7000/6781/.
14 Institute of Electrical and Electronics Engineers (IEEE), “IEEE Standard for Data Privacy Process,” IEEE 7002-2022, April 19, 2022,
https://standards.ieee.org/ieee/7002/6898/.
15 Louie Kangeter, “A Lifecycle Approach to AI Risk Reduction: Tackling the Risk of Malicious Use Amid Implications of Openness,”
Institute for Security and Technology, June 2024, https://securityandtechnology.org/wp-content/uploads/2024/06/A-Lifecycle-
Approach-to-AI-Risk-Reduction.pdf.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 6
 The AI Lifecycle Stages
The AI Lifecycle Framework breaks down the complex process of AI development into
manageable stages. This structured approach ensures a comprehensive understanding of
each phase, making it easier to target specific risk mitigation strategies effectively.

Data Collection & Preprocessing

Gathering raw data from various
sources and preparing it for use Model Architecture
in training the AI model; cleaning
Designing and structuring the
the data, handling missing values,
AI model, including the choice
normalizing data formats, and
of algorithms, network topology
augmenting datasets to enhance
(for neural networks), and other
their quality and relevance.
foundational elements that define
how the model processes data and
makes predictions.

Model Deployment
Integrating the trained AI model into a
production environment where it can
be accessed and used by end-users
or other systems. Includes setting up
Model Training & Evaluation
the necessary infrastructure, such as
Using preprocessed data to teach servers and APIs, to support model
the AI model to recognize patterns operation.
and make predictions; testing
the model’s performance using
separate validation datasets to User Interaction
ensure it generalizes well to new,
Designing the interfaces and
unseen data.
interactions through which end-users
engage with AI-powered applications.
This includes user experience (UX)
Model Application design, user interface (UI) design, and
Developing applications that use accessibility considerations.
the deployed AI models to perform
specific tasks. Includes designing
the software and systems that
leverage the AI model’s capabilities
for various user needs.

Ongoing Monitoring and

Maintenance
Continuously tracking the
performance of deployed AI models
and applications, addressing any
issues that arise, and updating
models as needed to ensure they
remain effective and secure.

March 2025 securityandtechnology.org 7

Return on Investment for
Implementing Strong Compliance
Practices
Non-compliance in building and deploying AI systems can result in consequences such as
reputational harm, erosion of public trust, and fines.16,17,18,19,20 Instead of being reactive, AI
builders and users can adopt proactive compliance practices that help accelerate and amplify
the value both builders and users can derive from the technology.21 There are various ways in
which strong compliance practices can generate ROI:

» Reduced regulatory risk exposure. Given the rapid proliferation of AI tools, industries
utilizing these technologies are expected to face increasing scrutiny from regulators.
Proactively implementing safety, security, privacy, transparency, and anti-bias
measures—and a compliance program to oversee their implementation—can help
prevent unexpected and costly harms, their associated litigation, and reputational
implications. For instance, in December 2024, just four compliance fines totaled up
to a hefty quarter billion euros for failing to comply with GDPR.22 Both GDPR and the
EU AI Act have extraterritorial reach, which means that some of the provisions apply
to companies that are not physically based in the EU but offer products and services
within the EU market. For instance, an AI lab based in the United States which makes
their AI tools available to EU users is subject to GDPR, EU AI Act, and other regulations
governing the EU market.

16 Elvira Pollina and Alvise Armellini, “Italy Fines OpenAI 15 Million Euros over Privacy Rules Breach,” Reuters, December 20, 2024,
https://www.reuters.com/technology/italy-fines-openai-15-million-euros-over-privacy-rules-breach-2024-12-20/.
17 Nikitha Anand, “The High Cost of Non-Compliance: Penalties Issued for AI under Existing Laws,” Holistic AI, March 28, 2024, https://
www.holisticai.com/blog/high-cost-non-compliance-penalties-under-ai-law.
18 Natasha Lomas, “MWC’s Organizer Slapped with GDPR Fine over Biometrics ID Checks Due Diligence,” TechCrunch, May 8, 2023,
https://techcrunch.com/2023/05/08/gsma-mwc-aedp-gdpr-dpia-fine/.
19 David Shepardson, “Lingo Telecom Agrees to $1 Million Fine over AI-Generated Biden
Robocalls,” Reuters, August 21, 2024, https://www.reuters.com/technology/artificial-intelligence/
lingo-telecom-agrees-1-million-fine-over-ai-generated-biden-robocalls-2024-08-21/.
20 CMS.Law, “GDPR Enforcement Tracker - List of GDPR Fines,” last accessed February 2025, https://www.enforcementtracker.
com/?insights.
21 Velu Sinha, Julie Coffman, Richard Fleming, Bill Groves, and Maria Teresa Tejada, “Adapting Your Organization for Responsible AI,”
Bain, January 2, 2024, https://www.bain.com/insights/adapting-your-organization-for-responsible-ai/.
22 CMS.Law, “GDPR Enforcement Tracker - List of GDPR Fines,” last accessed February 2025, https://www.enforcementtracker.
com/?insights.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 8
 » Competitive advantage. Strong compliance practices provide a competitive advantage
for both AI system builders and the enterprises adopting the systems. A recent report
from Bain reveals that organizations with an effective approach to responsible AI
doubled their profit impact from their AI efforts compared to those organizations that
lack such an approach.23

» Access to government procurement-shaped markets. The U.S. government’s

procurement policies and preferences make and shape markets. In 2023 alone, the
U.S. government invested more than $100 billion in information technology products
and services.24 A company that complies with the relevant standards in AI space will be
better prepared to compete in government procurement-shaped markets. For example,
logging features required to be turned on by default in government procurement-
shaped markets as a result of Executive Order 14028 on Improving the Nation’s
Cybersecurity then became industry standard for all cloud services in the market.
Additionally, as the AI market becomes one of the largest and one of the most valued
in the geopolitical and economic race to the top, governments will likely increase their
investments into the development of frontier models, likely favoring those companies
that have more robust security standards in place.25

» Ability to recruit and retain talent. Based on the working group members’ experiences
and observations, organizations that prioritize responsible AI development and
deployment practices have an edge in attracting top talent who increasingly seek
workplaces committed to responsible innovation. A strong ethical framework enhances
employee morale and loyalty, fostering an environment where skilled professionals
want to contribute and grow. This talent pipeline is crucial for both model capability
development as well as scaling AI products into new markets worldwide.

» Increased lifetime value. By investing in responsible AI practices, companies can build

stronger relationships with customers, partners, and employees, leading to higher
satisfaction and loyalty. For customers, this translates to increased lifetime value to
the company, as satisfied customers are more likely to return and advocate for the
brand, ultimately boosting long-term profitability. Proactively addressing AI compliance
concerns can safeguard an organization’s reputation over time. Companies that navigate
these challenges effectively are better positioned to withstand scrutiny and maintain
public trust, helping their brands remain resilient against potential controversies.

23 Velu Sinha et al., “Adapting Your Organization for Responsible AI.”

24 The White House, “Fact Sheet: OMB Issues Guidance to Advance the Responsible Acquisition of AI in Government,”
press release, October 2, 2024, https://bidenwhitehouse.archives.gov/omb/briefing-room/2024/10/03/
fact-sheet-omb-issues-guidance-to-advance-the-responsible-acquisition-of-ai-in-government/.
25 The White House, “Fact Sheet: OMB Issues Guidance.”

March 2025 securityandtechnology.org 9

 » Investor appeal. Enterprises that demonstrate compliance, particularly in emerging
technologies like AI, are likely to attract more investment, as stakeholders increasingly
consider security risks. A rigorous compliance program can indicate to investors that the
company has a lower risk threshold, prompting new investors, and sustaining existing
investors.26,27

Risk Mitigation Strategies for

Safeguarding Against Future
Failures
It is important to note that no risk management strategy, or combination of strategies, will
completely eliminate the possibility of an undesired outcome in any context. This is particularly
true in AI, as bad actors aggressively test the limits of their capabilities and as the potential for
“capability overhang”—defined as AI capabilities and aptitudes that were not envisioned by
their developers but emerge nonetheless—increases.28 As a result, it becomes challenging to
preempt unforeseen risks arising from novel capabilities and to foresee how malicious actors
could exploit them. As a rapidly developing frontier, AI has a limited track record from which
to design and implement effective controls. It also follows that an unintended consequence
should not always be attributed to a compliance failure, as not all AI risks and failure modes
are yet well understood. Such negative experiences can instead serve as learning trials for
builders, users, and regulators alike as they refine the state of practice in AI risk management.

The following table contains a selection of 22 technical and 17 policy-oriented risk mitigation
strategies co-created by the working group members and other contributors for alleviating,
pre-empting, or avoiding the three categories of compliance failure risks in the AI ecosystem.

26 Matthew White, Justin Daniels, and Javier Becerra, “AI Disclosures under the Spotlight: SEC
Expectations for Year-End Filings,” Baker Donelson, January 10, 2025, https://www.bakerdonelson.com/
ai-disclosures-under-the-spotlight-sec-expectations-for-year-end-filings.
27 Christopher Barlow, Brett Fleisher, David Simon, Nicola Kerr-Shaw, Melissa Muse, and Taylor Votek, “Rising Investment
in AI Requires Financial Sponsors to Address Unique Risks,” Skadden, Arps, Slate, Meagher & Flom LLP, January
14, 2025, https://www.skadden.com/insights/publications/2025/01/2025-insights-sections/the-deal-landscape/
rising-investment-in-ai-requires-financial-sponsors.
28 Zoë Brammer, “How Does Access Impact Risk?: Assessing AI Foundation Model Risk Along a Gradient of Access,” Institute for
Security and Technology, December 2023, https://securityandtechnology.org/wp-content/uploads/2023/12/How-Does-Access-
Impact-Risk-Assessing-AI-Foundation-Model-Risk-Along-A-Gradient-of-Access-Dec-2023.pdf.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 10
 Institutional failures
Lack of executive commitment to create a culture of compliance, establish necessary policies,
or empower success through the organizational structure (e.g., risk and audit board committees,
compliance officer role, quality assurance program), leading to foreseeable failures.

Procedural failures
Misalignments between an institution’s established policies as compared to its internal
procedures and staff training required to adhere to those policies.

Performance failures
An employee’s failure to follow an established process, or an automated system’s failure to
perform as intended, leading to an undesirable result.

We recognize that implementing all 39 of the below strategies may not always be feasible.
However, AI builders and users should consider which measures are appropriate according
to their context. This consideration should be proportional to factors such as the intended
use, potential risks, and application domain, which can range from entertainment and arts to
national security, healthcare, and finance.

Types of risks
Data Collection and Preprocessing (for builders) mitigated
1. Data collection requirements
Ensure that the collection, processing, and maintenance of personal or other protected
policy

data takes place in accordance with a valid legal basis. For instance, ensure that explicit
consent is obtained from individuals whose data is collected, with mechanisms to withdraw
consent at any point.

2. Privacy-preserving technologies
Protect sensitive data during the training stage by implementing privacy-preserving
technical

technologies—such as differential privacy and homomorphic encryption—during data

pre-processing so that, for example, the model does not learn personally identifiable
information. Implement data encryption both at rest and in transit to prevent label flipping
attacks and insecure data storage.29

29 Databricks, “Databricks AI Security Framework (DASF),” 2024, https://www.databricks.com/resources/whitepaper/

databricks-ai-security-framework-dasf.

March 2025 securityandtechnology.org 11

 3. Data source transparency through data cards
technical

For each model, publish a “data card” that documents the model’s data sources, privacy
measures, and preprocessing steps taken by its developers during the data collection and
model training phases.30,31,32,33

4. Bias detection tools for dataset auditing

Utilize automated bias detection tools to sift through training datasets and look for potential
technical

imbalances in attributes such as race, language, age, heritage, gender, viewpoint, etc.
Ensure that the training data is tested for accuracy and truthfulness to avoid negatively
influencing the model with non-factual information. Implement methods such as data
augmentation or re-weighting to mitigate potential biases.34,35

Types of risks
Model Architecture (for builders) mitigated
5. Cross-functional AI compliance team
Establish a cross-functional AI compliance team with representation from relevant
corporate functions such as legal, product, engineering, data infrastructure, cybersecurity,
policy

ethics, and internal audit functions. The team can blend together organizational strategies
at different lifecycle stages, harmonize internal policies and practices, and address
emerging issues related to compliance.
(Note, this mitigation applies to this and all subsequent lifecycle phases.)

6. Security program
Design or implement existing, reliable, robust cybersecurity and physical security controls
to secure model architecture and the infrastructure hosting the AI systems. Limit access to
policy

the system components to authorized personnel, with relevant aspects carefully managed,
controlled, and monitored.
(Note, this mitigation applies to this and all subsequent lifecycle phases.)

30 Nathalie Baracaldo and Hayim Shaul, “Fully Homomorphic Encryption,” IBM Research, February 9, 2021, https://research.ibm.com/topics/
fully-homomorphic-encryption.
31 Mahima Pushkarna, Andrew Zaldivar, Dan Nanas et al., “Data Cards Playbook,” People + AI Research, Google, March 5, 2021, https://sites.
research.google/datacardsplaybook/. According to Google’s “Data Card Playbook,” data cards are “structured summaries of essential facts
about various aspects of ML datasets needed by stakeholders across a project’s lifecycle for responsible AI development.”
32 Mahima Pushkarna, Andrew Zaldivar, and Oddur Kjartansson, “Data Cards: Purposeful and Transparent Dataset Documentation for
Responsible AI,” arXiv, April 3, 2022, https://doi.org/10.48550/arXiv.2204.01075.
33 “Regulation (EU) 2024/1689 EU Artificial Intelligence Act,” Official Journal of the European Union 2024/1689 (July 7, 2024), http://data.europa.
eu/eli/reg/2024/1689/oj.
34 Agnieszka Mikołajczyk-Bareła, Maria Ferlin, and Michał Grochowski, “Targeted Data Augmentation for Bias Mitigation,” arXiv, August 22,
2023, https://arxiv.org/abs/2308.11386.
35 “Pledge for a Trustworthy AI in the World of Work,” proceedings in the Summit for Action on Artificial Intelligence, Paris, February 11, 2025,
https://www.elysee.fr/emmanuel-macron/2025/02/11/pledge-for-a-trustworthy-ai-in-the-world-of-work.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 12
 7. Explainability by design
technical

Document and report an AI model’s features that explain its outputs, including the
contribution of specific training data points, while integrating explainability frameworks that
simplify complex machine learning models into easily understandable representations.36,37,38

8. Threat model-informed design requirements

technical

Simulate a variety of adversarial attacks to test and improve the robustness of the model
against malicious inputs to safeguard AI systems, especially in high-risk applications.39

9. Anomaly detection
technical

Incorporate anomaly detection and continuous monitoring mechanisms into model

architecture to identify unusual or malicious activity in real time and provide alerts for
potential misuse.40

10. Model cards

Create a model card for each user-facing model that documents its architecture, including
but not limited to performance metrics, explainability, safety measures, and robustness tests
technical

performed.41,42,43 Model cards can include documentation of the system’s intent, precise
scope (i.e., intended use cases and known limitations), as well as any “out of scope” uses
(i.e., what the model should not be used for) and the model’s known technical mitigations.
Update model cards periodically with newly observed model performance metrics, including
potential risks.

36 European Parliament, “EU AI Act: First Regulation on Artificial Intelligence,” European Parliament, June 8, 2023, https://www.europarl.europa.
eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence.
37 C3.AI, “LIME: Local Interpretable Model-Agnostic Explanations,” last accessed February 2025, https://c3.ai/glossary/data-science/
lime-local-interpretable-model-agnostic-explanations/.
38 Arize AI, “Explainability in Machine Learning: Top Techniques,” Arize Machine Learning Course, January 11, 2024, https://arize.com/
blog-course/explainability-techniques-shap/.
39 Jonas Rauber and Roland S. Zimmermann, “Welcome to Foolbox Native — Foolbox 3.3.3 Documentation,” Foolbox, 2021, https://foolbox.
readthedocs.io/en/stable/.
40 Louie Kangeter, “A Lifecycle Approach to AI Risk Reduction.”
41 Margaret Mitchell et al., “Model Cards for Model Reporting,” arXiv, January 14, 2019, https://arxiv.org/abs/1810.03993. According to Google’s
Model Cards Paper introduced in 2018, model cards “are short documents accompanying trained machine learning models that provide
benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups and intersectional
groups that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used,
details of the performance evaluation procedures, and other relevant information.”
42 OECD.AI, “OECD Framework for the Classification of AI Systems.”
43 NIST, “NIST AI RMF Playbook,” July 8, 2022, https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook.

March 2025 securityandtechnology.org 13

 Types of risks
Model Training and Evaluation (for builders) mitigated
11. AI safety benchmarks
Task AI compliance teams to start establishing mandatory safety benchmarks for
exceptionally capable models that stand to impact individuals and society in a highly
contextual fashion (e.g., contextualize based on use within specific industries or affecting
policy

vulnerable population groups).44,45,46 Integrate AI safety benchmarks, and specify that

models must pass certain safety criteria before deployment. Evaluate the models across
multiple axes, such as accuracy, fairness, bias, and robustness—akin to safety certifications
found in other industries (e.g., automotive crash tests).47,48 Specify that evaluations must be
conducted on diverse datasets to mitigate risks of overfitting when models are deployed.49

12. Benchmark hazard categories

Benchmark hazard categories (e.g. hate speech, Child Sexual Abuse Material (CSAM),
policy

violence, drugs, etc.) to guide training data selection and prompt generation. Create
labeled data with these safety categories in mind to improve how models classify and
identify risks.50

13. Model evaluation guidelines

Craft model evaluation guidelines to include metrics around algorithmic transparency,
policy

which would require the documentation of all training datasets, algorithm choices,
hyperparameter tuning, and metrics used to assess performance. These model evaluations
should be repeated periodically during training, especially for models that learn
continuously or adapt in real-time.51,52,53,54,55

44 European Parliament, “EU AI Act: First Regulation on Artificial Intelligence.”

45 Anthropic, “Anthropic’s Responsible Scaling Policy,” September 19, 2023, https://www.anthropic.com/news/
anthropics-responsible-scaling-policy.
46 Anca Dragan, Helen King, and Allan Dafoe, “Introducing the Frontier Safety Framework,” Google DeepMind, December 17, 2024, https://
deepmind.google/discover/blog/introducing-the-frontier-safety-framework.
47 MLCommons AI Safety Working Group, “Announcing a Benchmark to Improve AI Safety,” IEEE Spectrum, April 16, 2024, https://spectrum.ieee.
org/ai-safety-benchmark.
48 Insurance Institute for Highway Safety (IIHS) and Highway Loss Data Institute (HLDI), “Vehicle Ratings,” last accessed January 24, 2025,
https://www.iihs.org/ratings.
49 For example, after every epoch, developers can run safety benchmarks on the version of the model at hand to pinpoint emerging safety
deficiencies, and these benchmarking results could serve as data points to potentially adjust training objectives (e.g., reinforce guardrails
that help models avoid generating harmful responses).
50 “Pledge for a Trustworthy AI in the World of Work,” proceedings in the Summit for Action on Artificial Intelligence.”
51 International Standards Organization (ISO), “ISO/IEC DIS 42001,” 2023, https://www.iso.org/standard/81230.html.
52 NIST, “AI Risk Management Framework.”
53 OECD, “OECD Framework for the Classification of AI Systems,” OECD Publishing, no. 323 (February 2022), https://www.oecd.org/content/
dam/oecd/en/publications/reports/2022/02/oecd-framework-for-the-classification-of-ai-systems_336a8b57/cb6d9eca-en.pdf.
54 IEEE Standards Association, “IEEE Standard Model Process for Addressing Ethical Concerns during System Design,” IEEE 7000-2021,
September 15, 2021, https://standards.ieee.org/ieee/7000/6781/.
55 This component mirrors the UK framework’s mention of an iterative approach to risk management to address new risks as they come into
shape.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 14
 14. Data overfitting mitigations
technical

Guard against data overfitting, wherein a model performs well on training data but fails with
new, unseen prompts.56 Use out-of-distribution data to ensure models generalize well to
new prompts, rather than just performing well on benchmark-specific scenarios.57,58

15. Data provenance and watermarking

technical

Incorporate content provenance features in all model outputs, such as watermarks or

metadata that can verify the origin and integrity of generated content. This can prevent
bad actors from manipulating or misusing the model for harmful purposes and enhance
traceability.59,60

16. Bug bounty programs

technical

Create bug bounty programs to incentivize others to identify and report previously
unknown weaknesses in an AI model.61

17. Privacy-preserving technologies

technical

Ensure AI systems are privacy compliant by integrating privacy-preserving technologies

to minimize the danger of data exposure. This mitigation, and its performance, should be
included in any privacy compliance reports.62

18. Bias monitoring and data integrity checks

technical

Monitor potential biases during training through techniques such as adversarial debiasing.
Consider benchmarking model datasets on common fairness metrics such as demographic
parity and equalized odds to mitigate bias.63,64,65

56 IBM, “What Is Overfitting?” 2024, https://www.ibm.com/topics/overfitting.

57 To avoid data overfitting, developers can assess and compare the AI model’s performance on training versus test data, and track a number of
potential discrepancies in performance between both. If the performance of the model is significantly better on the training data, overfitting
may be the cause. Developers can also create an Out-Of-Distribution dataset, which comprises examples of data points not included in the
training set to measure the model’s performance on the OOD dataset and compare it with the model’s performance on its regular training set
to assess the model’s ability to generalize.
58 Alexandre Bonnet, “What Is Out-of-Distribution (OOD) Detection?” Encord, September 15, 2023, https://encord.com/blog/
what-is-out-of-distribution-ood-detection/.
59 Coalition for Content Provenance and Authenticity, “Guiding Principles,” 2024, https://c2pa.org/principles/.
60 Kizuna, “The Hiroshima AI Process: Leading the Global Challenge to Shape Inclusive Governance for Generative AI,” The Government of
Japan, February 9, 2024, https://www.japan.go.jp/kizuna/2024/02/hiroshima_ai_process.html.
61 Louie Kangeter, “A Lifecycle Approach to AI Risk Reduction.”
62 The White House, “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial
Intelligence,” October 30, 2023, https://bidenwhitehouse.archives.gov/briefing-room/presidential-actions/2025/01/14/
executive-order-on-advancing-united-states-leadership-in-artificial-intelligence-infrastructure/.
63 IBM, “AI Fairness 360,” aif360.res.ibm.com, accessed January 24, 2025, https://aif360.res.ibm.com/.
64 Jenny Yang et al., “An Adversarial Training Framework for Mitigating Algorithmic Biases in Clinical Machine Learning,” Npj Digital Medicine 6,
no. 1 (March 29, 2023), https://doi.org/10.1038/s41746-023-00805-y.
65 IGI Global, “What Is Equalized Odds,” 2023, https://www.igi-global.com/dictionary/fairness-challenges-in-artificial-intelligence/115386.

March 2025 securityandtechnology.org 15

 19. Secure training pipelines
Train models in a secure environment with version control and cryptographic measures to
technical

prevent unauthorized changes to datasets or model parameters. Apply penetration testing

on AI training environments to identify and address vulnerabilities that could be exploited
for malicious purposes. Record model performance and evaluation metrics in model cards
for future auditing.

Types of risks
Model Deployment (for builders and users) mitigated
20. Incident reporting and disclosure framework
Develop an incident reporting and response framework that requires AI system breaches
policy

and incidents to be documented and tracked. Include steps to escalate and report
violations, such as jailbreaking.66,67 This framework could be leveraged in periodic or ad hoc
reporting to an organization’s compliance team, particularly for new tools being developed
or deployed.68

21. Staff training

Design and implement mandatory compliance training for staff members involved in the AI
supply chain. Training modules should be role-specific and take into account geographic
policy

jurisdiction and use context. All staff members utilizing AI tools should also demonstrate
minimum literacy of AI system functions and limitations, intended use, and potential
impact.69,70
(Note, this mitigation applies to this and all subsequent lifecycle phases.)

22. Deployment plan

Consider deploying the system following a predefined, approved plan that outlines the AI
system’s inventory, maintenance, roles of involved actors, timeline, and a context-specific
policy

testing and feedback strategy aligned with the model’s risk profile. The plan should also
account for resource issues such as memory, compute, network, storage, redundancy,
and load balancing. It should define risk thresholds, and incorporate digital, physical, and
environmental security procedures to safeguard system assets.71

66 Thorn and All Tech Is Human, “Safety by Design for Generative AI: Preventing Child Sexual Abuse,” Thorn Repository, 2024, https://info.thorn.
org/hubfs/thorn-safety-by-design-for-generative-AI.pdf.
67 Zeqiu Wu et al., “Fine-Grained Human Feedback Gives Better Rewards for Language Model Training,” arXiv, October 30, 2023, https://doi.
org/10.48550/arXiv.2306.01693.
68 Sean McGregor et al., “To Err Is AI: A Case Study Informing LLM Flaw Reporting Practices,” arXiv, October 15, 2024, https://arxiv.org/
pdf/2410.12104.
69 European Commission, “First Rules of the Artificial Intelligence Act Are Now Applicable,” Shaping Europe’s Digital Future, 2025, https://
digital-strategy.ec.europa.eu/en/news/first-rules-artificial-intelligence-act-are-now-applicable.
70 Oliver Yaros et al., “EU AI Act: Ban on Certain AI Practices and Requirements for AI Literacy Come into
Effect,” Mayer Brown LLP, January 31, 2025, https://www.mayerbrown.com/en/insights/publications/2025/01/
eu-ai-act-ban-on-certain-ai-practices-and-requirements-for-ai-literacy-come-into-effect.
71 UK Government, “National AI Strategy.”

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 16
 23. Transparency measures
Document and publicize (as appropriate) comparisons of a new AI model with existing
technical

models, infrastructure and tools, data accessibility, accuracy, interpretability, complexity,

training time, and scalability. Implement model disclosure frameworks that include
extended model cards, automated verification with reproducible testing and validation
mechanisms, an adjudication process to fairly assess models, as well as a dynamic scope
for models to adapt to emerging common uses.72,73

24. System integration

Integrate AI models into existing technical architectures and legacy systems to ensure they
technical

are practicable, accessible, and user-centric on both the back-end and front-end. Ensure
that system integration processes account for compatibility with legacy systems, potential
performance degradation, and potential data integration challenges.74 Consider first testing
in a sandbox to discover compatibility issues prior to integration.

Types of risks
Model Application (for builders and users) mitigated
25. Application-specific security controls
policy

When designing or deploying a specific AI tool, consider creating a decision tree to help
choose which AI tool to deploy.75 The decision tree should differ for AI tools used internally
versus those used for business-to-user or business-to-business interactions.

26. Query rate limits

technical

Set a limit on the number of queries a user can input into an AI model within a specific
timeframe to mitigate AI model abuse, including through automated means.76,77

27. Human in the loop

Mandate the inclusion of human oversight and control mechanisms in AI applications,
technical

especially for high-risk or sensitive use cases, to prevent fully autonomous unsanctioned
actions. Define specific use cases in which agentic AI capabilities will provide operational
advantages (e.g., increase productivity or efficiency) and cases in which keeping the human
in the loop is essential for taking specific actions. Implement appropriate human-feedback
loops and checks to assess the AI decision-making process and intervene when needed.

72 Sven Cattell, Avijit Gosh, and Lucie-Aimée Kaffee, “View of Coordinated Flaw Disclosure for AI: Beyond Security Vulnerabilities,” Proceedings
of the AAAI/ACM Conference on AI, Ethics, and Society 7, no. 1 (2024), https://doi.org/10.1609/aies.v7i1.31635.
73 Anthropic, “Anthropic’s Transparency Hub,” last updated February 27, 2025, https://www.anthropic.com/transparency.
74 NIST, “AI Risk Management Framework.”
75 U.S. Department of Energy, “Cybersecurity Considerations for Procurement,” Federal Energy Management Program, October 2024, https://
www.energy.gov/femp/cybersecurity-considerations-procurement.
76 OpenAI, “OpenAI O1 and O1-Mini Usage Limits on ChatGPT and the API,” 2025, https://help.openai.com/en/
articles/9824962-openai-o1-preview-and-o1-mini-usage-limits-on-chatgpt-and-the-api.
77 Anthropic, “Rate Limits,” last accessed February 2025, https://docs.anthropic.com/en/api/rate-limits.

March 2025 securityandtechnology.org 17

 Types of risks
User Interaction (for builders and users) mitigated
28. User consent
Develop policies to ensure users are informed prior to an AI system making a decision on
policy

their behalf. For systems supporting high-impact use cases such as employment, financial,
or healthcare decisions, provide users with clear explanations (using model cards or other
techniques) of how decisions are made and how to appeal them. Ensure that user-AI
interactions are governed by clear user consent mechanisms.78

29. Robust user feedback loops

policy

Integrate mechanisms for users to provide feedback or contest decisions made by the AI
system, to protect user autonomy and promote ethical engagement.79

30. Education programs for end-users

policy

Implement programs to educate end-users about the limitations and proper use of an AI
model, including safety measures to consider while interacting with the model. This would
potentially increase public trust in AI by promoting informed interactions.

31. “Opt out” option for end-users

Provide an explicit option for users to ‘opt out’ of processes in which decisions are made
technical

automatically by AI models and provide the option for human operators to be involved
instead. Ensure that users are notified when an AI system is involved in generating content,
advice, decisions, or actions and are provided with clear explanations of the criteria behind
these outcomes.80,81

32. Watermarking techniques

technical

Adopt watermarking techniques to identify AI-generated outputs for users’ awareness.82

While watermarking is not a holistic solution and can be vulnerable to tampering, it is a
preliminary step to help users distinguish between traditionally produced and AI-generated
content.

78 International Standards Organization (ISO), “ISO/IEC DIS 42001,” ISO, 2023, https://www.iso.org/standard/81230.html.
79 International Standards Organization (ISO), “ISO/IEC DIS 42001.”
80 “Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with
regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data
Protection Regulation),” Official Journal of the European Union 119/1 (May 4, 2016), https://eur-lex.europa.eu/legal-content/EN/TXT/
PDF/?uri=CELEX:32016R0679.
81 Institute of Electrical and Electronics Engineers (IEEE), “IEEE Standard Model Process for Addressing Ethical Concerns during System Design,”
IEEE 7000-2021, September 15, 2021, https://standards.ieee.org/ieee/7000/6781/.
82 Restack, “Watermarking Techniques in AI” Restack.io, 2025, https://www.restack.io/p/ai-in-iot-answer-watermarking-techniques-cat-ai.

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 18
 Ongoing Monitoring and Maintenance (for builders and Types of risks
users) mitigated

33. AI compliance reviews

policy

Task the AI Compliance Team to conduct periodic reviews during which models are audited
to ensure continued alignment with relevant regulations, frameworks, and internal policies.
Document and update all audits in the model cards to maintain transparency.83,84,85,86

34. Responsible information sharing

policy

Uphold clear processes for responsibly sharing AI safety and security information with
relevant stakeholders (i.e., governments, industry, civil society), to include security risks,
potential vulnerabilities, and ways to mitigate misuse.87

35. System transition and decommission

Ensure that the AI system adheres to a transition or decommissioning plan that complies
policy

with applicable laws and regulations, protecting users’ privacy and data rights, disposing
of sensitive materials, and retaining system documentation for developers and the
organization.

36. Third-party reviews

Integrate periodic independent reviews to assess an AI model against safety, security,
policy

and performance quality metrics. These reviews could also include pre-deployment risk
assessments and can be informed by insights from AI governance and policy-focused
organizations.88

37. Monitoring for model drift

technical

Use automated monitoring systems to track model performance over time and detect
model drift or data drift. Implement mechanisms that can be triggered in the event a model
starts behaving unpredictably, which might lead to humans retraining it.

83 National Institute of Standards and Technology, “AI Risk Management Framework.”

84 “Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with
regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data
Protection Regulation),” Official Journal of the European Union 119/1 (May 4, 2016), https://eur-lex.europa.eu/legal-content/EN/TXT/
PDF/?uri=CELEX:32016R0679.
85 European Parliament, “EU AI Act: First Regulation on Artificial Intelligence,” European Parliament, June 8, 2023, https://www.europarl.europa.
eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence.
86 Reto Grubenmann and Flavia Masoni, “ISO/IEC 42001: The Latest AI Management System Standard,” KPMG, accessed January 26, 2025,
https://kpmg.com/ch/en/insights/technology/artificial-intelligence-iso-iec-42001.html.
87 G7, “Hiroshima Process International Guiding Principles for Organizations Developing Advanced AI Systems.”
88 Monika Viktorova and Hadassah Drukarch, “Operationalizing Independent Review in AI Governance,” Responsible AI, November 25, 2024,
https://www.responsible.ai/operationalizing-independent-review-in-ai-governance/.

March 2025 securityandtechnology.org 19

 38. Model termination guidelines
technical

Develop clear emergency response protocols that specify under what circumstances an AI
system would immediately be shut down, how this process would be carried out, and how it
can be verified.

39. Monitoring protocols and logging

technical

Ensure that AI systems are designed to log all operational activities and AI-generated
outputs such as reports, predictions, recommendations, and to provide the relevant
stakeholders access to the recorded information.89,90

Conclusion
Charting a path towards effective AI compliance measures requires the coordinated efforts
of diverse stakeholders throughout the AI ecosystem. While this paper offers actionable risk
mitigation strategies that AI builders and users can implement, there remains a need for broader
collaboration on AI compliance. Safeguarding against future failures in the AI ecosystem
requires a multidisciplinary approach; a technology sector that has a potential and ambition for
universality should take insights from a broader array of stakeholders, including philosophers,
ethicists, anthropologists, linguists, psychologists, and practitioners in human-computer
interaction, user experience, and other disciplines.

AI ecosystem stakeholders should accelerate information sharing among verified researchers

from leading labs, universities, and startups to diffuse best practices and methods for responsible
AI development. By doing so, AI ecosystem stakeholders will have access to continuous learning
resources, and AI builders and users will not have to choose between innovating and being
responsible.

89 European Parliament, “EU AI Act: First Regulation on Artificial Intelligence.”

90 NIST, “AI Risk Management Framework.”

Navigating AI Compliance, Part 2: Risk Mitigation Strategies for Safeguarding Against Future Failures 20
INSTITUTE FOR SECURITY AND TECHNOLOGY
www.securityandtechnology.org

[email protected]
Copyright 2025, The Institute for Security and Technology

March 2025 securityandtechnology.org 21