In Parts I and II, we examined ten corporate compliance risks arising from Anthropic’s Registration Statement. In Part III, we translated those risks into five lessons for compliance professionals. Today, I want to explore 12 questions where those lessons should land at the board level. Anthropic’s Safety and Risk Committee (the “Committee”) provides a specific place to begin.
The registration statement describes a committee consisting of two independent directors, Christopher Liddell and Vas Narasimhan, with Narasimhan serving as chair. Its responsibilities encompass oversight of model safety and security, including the policies, processes, and controls for managing significant model risks. The committee also assists the board in overseeing other material risks, including legal, regulatory, cybersecurity, privacy, and compliance risks.
That breadth matters. Model safety, commercial decisions, and enterprise compliance can affect one another. A new capability can change misuse exposure. A distribution partnership can complicate access restrictions. A deployment decision can create questions about whether customer representations remain supportable.
The filing states that the committee operates under a board-approved written charter. Its summary does not establish every approval right, reporting arrangement, or escalation threshold. The following twelve questions are my recommended oversight agenda, rather than a description of procedures Anthropic has adopted or findings that its controls are deficient.
1. Which risks require the committee’s closest attention?
A broad mandate requires disciplined prioritization. The Committee should ask management which combinations of model capability, user access, and business activity could create the most serious harm or compliance exposure.
What has changed since the last review? Has a model gained new capabilities, an agent received broader permissions, or a commercial arrangement introduced new users and jurisdictions? Which changes require additional safeguards before deployment or expansion?
Evidence to request: A prioritized assessment connecting material risks to accountable executives, key controls, unresolved gaps, and decisions requiring committee attention. The assessment should explain uncertainty and the risk remaining after safeguards operate.
The practical lesson is to organize the agenda around consequential decisions. A catalogue of risks becomes useful when directors can see what management must do about them.
2. Are responsibilities and escalation routes clear?
The Committee should ask how its work connects to the full board and other Committees. The filing also assigns risk assessment and risk management policy review to the audit committee. Overlapping responsibilities make coordination essential.
Who leads when a safety incident raises questions about investor disclosures, financial exposure, or executive incentives? What requires immediate notification to the committee chair? When must an unresolved issue reach the full board?
Evidence to request: A responsibility map, notification thresholds, and records demonstrating how significant matters move between management, committees, and the board.
The Committee should also establish what happens when management recommends deployment despite a serious unresolved concern. Directors need clarity about who decides, who advises, and who receives notice. They should confirm any authority to require a pause against the actual charter and board delegations.
3. Can independent challenge reach the committee?
Technical expertise is essential, but expertise alone does not guarantee that dissent reaches directors. The Committee should ask whether compliance, legal, security, and safety leaders can present concerns directly, including disagreements with senior management.
Where do staffing, specialist expertise, or restrictions on information limit effective challenge? Which recommendations have management declined, and what exposure remains? Can the Committee obtain independent advice when internal assessments conflict?
Evidence to request: Reporting arrangements, material resource gaps, significant unresolved disagreements, and the process for obtaining outside expertise.
Private sessions with relevant control leaders can help directors understand what formal presentations omit. Those sessions should produce clear follow-up where necessary. The test is whether an uncomfortable finding can reach the committee early enough to influence a consequential decision.
4. Do incentives and culture support safety commitments?
The committee should examine how commercial and development pressures affect employee behavior. What happens when testing identifies a concern that could delay an important release? Are employees rewarded for surfacing weaknesses and resolving them?
How does management identify retaliation or other consequences for raising concerns? Do employees have reporting channels that work when an issue crosses technical, legal, and commercial boundaries?
Evidence to request: Employee feedback, significant safety and compliance reports, retaliation follow-up, and examples of decisions changed because employees challenged an assumption.
Where incentives appear to undermine controls, the Committee should coordinate with other relevant committees, such as a compensation committee (if one is ever created). A falling report count requires explanation. It could reflect fewer problems, less trust in reporting, or weaker detection. Directors should ask what evidence supports management’s interpretation.
5. Do safety controls work under realistic conditions?
The Committee should ask what testing establishes about the effectiveness of important safeguards. Which scenarios were examined, what failed, and what limitations remain in the assurance provided?
A model performing well in a controlled evaluation may behave differently when connected to tools, sensitive information, or persistent workflows. Directors should ask whether testing reflects the deployment environment and foreseeable attempts to defeat restrictions.
Evidence to request: Evaluation coverage, significant failures, independent challenges, exceptions, and evidence that corrective measures have been retested.
A useful question is: Show us an instance in which a safeguard detected a serious problem and enabled intervention. Management should explain the detection, the response, and what the example does and does not establish about broader control effectiveness.
6. Are autonomous agents constrained by enforceable limits?
Agent oversight deserves specific attention because authority to act changes the consequences of error or misuse. The Committee should ask what agents can access, change, transmit, purchase, or execute without additional approval.
What prevents an agent from exceeding its permissions? Which actions require meaningful human review? Can the company revoke credentials, contain activity, and preserve evidence when behavior becomes unsafe?
Evidence to request: Permission boundaries, approval requirements, monitoring results, intervention exercises, and records of unauthorized or unexpected actions.
Directors should scrutinize the human review process. A reviewer needs sufficient information, time, and authority to challenge the proposed action. Management should demonstrate that intervention works under realistic conditions and that safeguards remain effective after changes to tools or permissions.
7. Are data protection and intellectual property risks connected to safety oversight?
The Committee’s enterprise remit should bring data and intellectual property questions into the discussion. Which sensitive information supports model development or enters deployed systems? What access, retention, and use restrictions apply?
Can management substantiate its rights to use critical datasets and software? How would the business respond if those rights were challenged or access were lost? What could a security compromise expose beyond immediate service disruption?
Evidence to request: Material data flows, significant security findings, unresolved rights questions, access controls, and incident-exercise results.
The Committee should examine operational consequences alongside legal exposure. A disputed dataset, compromised credential, or exposed model asset may require changes to products, access arrangements, and customer communications. Management should explain how those decisions would be coordinated.
8. Can partners preserve the safeguards Anthropic relies upon?
The Committee should ask where cloud providers, distribution partners, data suppliers, and other critical relationships affect the operation of safety and compliance controls.
Which safeguards depend on a partner’s actions? Can management detect prohibited access or misuse through indirect channels? What information, notification rights, and assurance evidence are available when a partner controls part of the environment?
Evidence to request: Critical dependency maps, material contract exceptions, unresolved diligence findings, monitoring results, and tested contingency plans.
Concentration requires particular scrutiny. Several commercial relationships may depend on the same infrastructure. The Committee should ask what management could actually do if a critical provider failed to meet requirements, including how quickly access could be restricted or essential services maintained.
9. Do legal restrictions translate into operational controls?
Sanctions, export controls, anti-corruption obligations, and government interactions require attention across the commercial lifecycle. The Committee should ask how management identifies relevant restrictions and converts them into controls over sales, delivery, support, and access.
Where could resellers, shared accounts, intermediaries, or unusual payment arrangements obscure the actual customer or activity? Which exceptions require escalation, and who approves them?
Evidence to request: Material screening exceptions, access-control testing, high-risk intermediary reviews, and significant unresolved compliance matters.
The Committee should connect these findings to deployment decisions. A policy prohibiting certain access provides limited assurance if the delivery architecture cannot enforce the restriction or identify attempts to bypass it. Management should explain both preventive controls and detection capabilities.
10. Are external claims consistent with internal findings?
The Committee should examine how evidence about safety, reliability, privacy, and security informs product representations and material disclosures. Which claims depend on testing with significant limitations?
What happens when an incident contradicts a public statement? How do safety and compliance findings reach legal, communications, finance, and the disclosure process?
Evidence to request: Substantiation for significant claims, unresolved discrepancies, and records showing how material incidents affect communications decisions.
Financial reporting and investor disclosure matters require coordination with the audit committee and responsible management. The Safety and Risk Committee’s contribution is to ensure that relevant risk information reaches those processes. Ask management which important claim would be hardest to substantiate today, and why.
11. Are investigations independent and connected to decisions?
A significant incident may involve technical failure, misconduct, or both. The Committee should ask who determines investigative scope when allegations concern senior executives, deployment decisions, or the suppression of safety findings.
Are evidence preservation and interim safeguards timely? Who evaluates reporting obligations and potential disclosure decisions with appropriate counsel? What requires independent oversight?
Evidence to request: Suitably protected summaries of significant matters, independence arrangements, interim measures, and outstanding decisions.
The Committee should also examine how the issue came to light. Internal detection, an employee report, a customer complaint, and external discovery reveal different strengths and weaknesses. That detection path should inform improvements to monitoring and reporting.
12. Does remediation withstand independent verification?
Closing an action item should require evidence that the underlying problem was addressed. What failed in the process, permissions, incentives, or supervision? Where else could the same weakness exist?
Who owns each corrective action, what remains overdue, and which interim safeguards protect the business while work continues?
Evidence to request: A remediation record showing severity, owner, deadline, verification results, and repeat findings.
Retesting should reflect the conditions that produced the original failure. The committee should ask who performed the validation and what limitations remain. A revised policy or completed software change is an output. Demonstrated improvement in control performance is the outcome directors need to understand.
An agenda for the next committee meeting
Ask management to present the five most consequential unresolved safety and enterprise compliance risks. For each, require the evidence, accountable executive, available options, and decision needed. Schedule a private session with relevant control leaders and examine one recently closed remediation item.
Confirm coordination arrangements with the audit and compensation committees. Record substantive challenges, decisions, responsibilities, and follow-up. The final question should be practical: What must this committee escalate or recommend today to make Anthropic’s safety and compliance commitments effective in operation?
