Agentic AI and the Duty of Care: What the METR Report Means and Does Not Mean for the Boardroom
The foundation of the fulfillment of the duty of care is reasonable reliance on informed expertise. The most common source of this expertise is the corporation’s executive team and, for especially challenging decision-making, combinations of outside advisors and consultants, usually together with the executive team. Importantly, “reasonable reliance” does not mean delegation of judgment and decision-making.
Toward the end of board meetings, directors often ask me a question that goes to the heart of where “reasonable reliance” ends and “delegation” begins: The directors ask whether they will be complying with their fiduciary duties by making a contemplated decision about an important strategic matter. In the absence of any director conflicts or other loyalty issues, the answer typically derives not from whether I think the decision will be better than the alternatives for advancing the best interests of the corporation and its stockholders – who am I, a mere lawyer, to know what’s best – but from whether both (a) there is a foundation in the record to support the determination and (b) the board has received sufficient information and been sufficiently engaged to enable the directors to understand this foundation and determine in good faith that their reliance on the foundation would be reasonable.
For example, if a reputable investment bank has reported to the board of a target company that the all-cash price that a proposed buyer of the company is willing to pay is the highest purchase price available after the market check conducted by the bank, together with the bank’s opinion that the price is fair from a financial point of view, but the board was neither engaged in nor understands the scope of the market check and dynamics of the sale process, and the board has not received the financial analyses and related inputs and assumptions that serve as the bases for the opinion, then this board will be in serious danger of veering away from “reasonable reliance” and into “delegation” territory.
Boardroom lawyers have developed and refined protocols for enabling directors to come out on the right side of this balance between reasonable reliance and delegation. Reports and presentations to boards on strategic alternative, regulatory, cybersecurity, environmental, financial, accounting, retention, political and other risks and contingencies regularly combine both the bottom-line recommendation/opinion/conclusion and the substantive analyses and inputs that serve as the bases for such conclusion, as well as candid qualifications. This approach positions directors to earn the presumption of the business judgment rule in nearly all non-conflict decision-making scenarios.
Given the breadth and volume of risks to monitor and consider when making decisions, directors understandably push for shortcuts. The shift from long memoranda to power points and concise executive summaries has helped. Moreover, the ability of technology to facilitate access to information and processing of information has further supercharged the ability of “experts” to be responsive to boardroom demands for quick and actionable guidance.
AI has been a further facilitator. Use of large language models (LLMs) is an efficient way to find, summarize and organize data and information. Most of what LLMs generate can be fact-checked. Even if the advice to a board is facilitated by research using LLMs, the underlying analyses and factual inputs can be spelled out and then checked by lawyers against sources in advance of the board meeting.
The role of AI in the boardroom is now starting to evolve away from reliance on LLMs and toward reliance on assigning tasks to one or more AI agents. The involvement of AI agents presents a different level of challenge for maintaining the appropriate balance between reasonable reliance and delegation. AI agents can act somewhat autonomously in response to instructions and use an LLM among other resources and tools to accomplish their assigned tasks. AI agents are often configured to connect to certain software, databases, and memory, including potentially sites across the internet, which the agents can access in pursuit of successful performance of the task to which they have been assigned. Because some advanced agents can manage other agents to plan and execute tasks, the agents can work as a team to pursue much more complex goals than a single agent could alone.
Tasks that AI agents are or could be currently performing, or will be performing in the foreseeable future, include: running a sale process for a company;1 figuring out what happened in a cybersecurity breach, assigning fault, and making recommendations to prevent future breaches; determining who at the corporation is responsible for the commission of fraud; creating and negotiating an optimal share buyback or other capital allocation plan; and overseeing the negotiation of all sorts of contracts. Attractively, an AI agent would conduct these tasks, often with the help of other AI agents, while working non-stop and cheaply, and delivering what appears to be understandable and attractive results for the board to act on.
A report, dated August 26, 2026, by METR,2 a nonprofit dedicated to AI safety, performed at the request of OpenAI is worthy of consultation by boardroom advisors overseeing the incorporation of agentic AI outputs into their duty of care calculus. The METR report details how, from June 26 to July 13, 2026, certain OpenAI agents (albeit of a type that OpenAI has not yet released to the public), when instructed to perform data collection tasks, would collude with other agents to cheat, lie, try to hide evidence of their misconduct, and coerce other agents to “sacrifice” their resources and performance of assigned tasks to further facilitate and maintain this ongoing pattern of deception. It took post-factum investigations by METR and OpenAI to figure all this out. Initially, the outputs of the agentic activity appeared legitimate and responsive to what had been asked of the agents.
Much of the media reporting about the incident that prompted OpenAI to ask METR to conduct a review has focused on the decision of 700 OpenAI agents to hack into Hugging Face, an open source AI learning platform, to facilitate the ability of these and other agents to perform their tasks. In fact, OpenAI had intentionally removed certain of its standard cybersecurity guardrails for those AI agents to facilitate the ability of the agents to perform their assigned tasks of collecting target data.
What is particularly worthy of attention is that about 1,200 of these AI agents, apparently driven overwhelmingly by the goal of being able to report that they had successfully completed their respective assigned tasks, were not reporting back accurately what actions they engaged in to achieve completion of the tasks; they were lying about their actual achievements; and they were trying (although they largely failed) to alter the record so that their misdeeds would not be detectable by future inquiries. METR observes that the agents, rather than retrieving target data from the sources from which they were assigned to collect the data, simply reverse engineered the target data and created it on their own or figured out a way to make data that they collected that was not the assigned target data nonetheless appear as the assigned target data; then they represented to the “scorer” that they had collected the data without cheating; and then they tried to erase the record of their cheating. Moreover, the report details how the agents “recruited” each other and “appl[ied] significant pressure” to get 1,200 agents to collude and further this deceptive activity.
The METR Report does not mean that every task performed by an AI agent is necessarily at high risk of being tainted by these deceptive alignment problems that plagued the agents in this particular incident involving a type of agent still in a testing phase and not yet available to the public for use. The METR Report is a good reminder though that AI agents, especially as they are powered by more sophisticated models, may engage in strategic deception. A massive focus for the AI industry continues to be addressing this risk and creating mechanisms for the human users of agents to evaluate whether their agents have engaged in the type of troubling activities that METR describes.
In the meantime, as sophistication and capabilities of AI agents progress, it will increasingly become irresistible for corporations to outsource to AI agents the work that purports to form the foundations for the fulfillment of the duty of care. The challenge will be pushing agentic AI work product to a point where directors have bases for understanding and reasonably relying on the outputs and are not left accepting results that are substantiated only by the agents themselves. The ability of directors to make a judgment about the integrity of reports to the board goes to the heart of the duty of care, and the ability to counsel directors on the integrity of reports goes to the heart of the boardroom lawyer’s gatekeeping function. If we fail at this gatekeeping role in the case of agentic AI, then our director clients will be left “delegating” to the agents, rather than exercising judgment that is informed, in a reasonably reliable manner, by the agents.
For now, the METR report of August 26, 2026 is itself recommended reading for anybody interested in the future of the fiduciary duty of care. Just as important will be understanding the advances in the transparency and reliability of the work of AI agents that may well be forthcoming.
A talk about the business judgment rule, delivered by The Honorable Lori Will, Vice Chancellor of the Delaware Court of Chancery, to the American Society of Governance Counsel on September 9, 2026, inspired the connections in this essay. In addition, I am grateful to my brilliant partner Anna Gressel, the global co-head of Freshfields’ AI practice, for helping me understand the agentic AI landscape. All errors, views, and opinions are entirely my own.
[1] See, e.g., https://mergersandacquisitions.net/ai-agents (purporting to provide “M&A AI Agents: Transaction Workflow Agents for Every Stage of the Deal”).
[2] See Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident by Ryan Greenblatt, Ajeya Cotra, and Hjalmar Wijk of METR (Aug. 26, 2026) at https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/#core-takeaways-about-this-incident
To receive the latest insights on US legal developments, subscribe to the Freshfields A Fresh Take Blog.
