Wachtell Lipton wants boards to know what their own AI agents are allowed to do

Delaware makes oversight liability hard to establish. Under the Caremark line of cases, directors face it only where they utterly fail to implement a board-level reporting system, or consciously fail to monitor it or respond to red flags ^[5]. The claim also needs a showing of bad faith ^[5]. Martin Lipton, Kevin S. Schwartz and David M. Adlerstein, the memo's authors ^[1], wrote that "ineffective or unsuccessful oversight alone is not enough" ^[6]. Courts have still let stockholder claims proceed to discovery where the complaint alleges those failures with specificity ^[7].

The memo's main request is a test boards apply to their own structures. Directors "should consider whether existing reporting-and-escalation structures adequately capture AI-related cyber risks," the authors wrote ^[8]. Caremark asks whether a reporting system exists and whether anyone watched it ^[5]. A reporting line built before a company let software agents into its data can pass the first part of that test and still say little about the agents.

For agents, the board's share is a general understanding of the actions they are permitted to take, the controls on their reach into systems and data, the testing of those controls, and the points where a significant action needs human approval ^[11]. Management builds the rest. That includes a framework identifying the company's most critical data, where it resides and how AI systems use it, with clear accountability for protecting it ^[9]. It also includes assessing key AI vendors, so the board can see where the company depends on them and what contingency plans exist ^[12]. Training covers personnel at every level of seniority and may include periodic simulations for spotting deepfakes ^[14].

An audit committee chair could fairly say the quarterly cyber briefing already covers this. The memo largely keeps that arrangement. Management tracks AI-enabled threats and defenses, drawing on the NIST Cybersecurity Framework and qualified outside experts, and the board receives periodic reports on significant incidents, testing results, remediation status and whether mitigation is working ^[10]. The difference is in what the briefing has to contain, scaled to a company's size, industry, regulatory environment, data assets, AI uses and threat profile ^[13].

The loss figures are smaller per incident than the capability claims suggest. Reported losses nearing $900 million across the FBI's AI-tagged complaints for 2025 ^[4] come to less than about $41,000 per complaint on average ^[15]. The memo's case rests on capability: AI systems that find and exploit latent vulnerabilities at increasingly modest cost, at a speed and scale beyond human hackers or conventional tools ^[2], plus the April and July incidents ^[17][18]. The memo does not identify the model or the company involved.

I think the order of work matters more than the list. The capability claims rest on a handful of incidents, while the reporting question is one a board can answer this quarter ^[8]. An inventory of agent permissions and a data map drawn up now become the baseline that later testing results and remediation reports are read against ^[9][11]. Without it, the periodic reports the memo describes have nothing to be compared with ^[10].