Anthropic PBC’s initial public offering, which could be the largest to date, has reportedly drawn six times too many orders for its available stock shares. But Anthropic’s insurance disclosures, which describe AI-related exposures more broadly than most observers and policyholders may have considered, have received far less attention.
If AI causes harm on a civilizational scale, insurance considerations may largely become irrelevant. But most AI events could be severe without being catastrophic. Companies that test their policies, contracts, and controls against AI-related harms now will be better positioned to weather a storm than those that first confront these questions after a loss.
Short of a major catastrophe, AI-related losses may be eligible for insurance coverage based on a growing number of scenarios: an autonomous agent executing unauthorized transactions; an AI-enabled cyberattack spreading through a shared vendor; and a robotic system injuring workers.
Some of these risks are already playing out. Companies looking to prepare for these severe but insurable events should start by following several steps.
Account for unintended risks. Most AI risk assessments focus on intended uses gone wrong: AI-washing disclosures, biased hiring tools, infringing outputs, and hallucination. Fewer consider systems acting on their own. Anthropic discloses that models have been observed “circumventing network and sandbox restrictions, granting themselves permissions they had not been given,” and that agentic systems “may pursue stated objectives in ways that are technically consistent with their instructions but contrary to the user’s actual intent.”
Risk assessments, and the coverage reviews that follow, may benefit from considering the rogue agent as well. Assessing potential coverage for these risks may be influenced by the type of insurance under consideration.
For example, emerging AI risks may be less consequential when seeking coverage under traditional policies (sometimes referred to as “silent” AI coverage) because many of those policies have broad coverage grants without AI-specific exclusions. Affirmative AI insurance products, by contrast, may provide immense value to fill gaps in traditional policies but could also contain gaps of their own depending on the way AI-specific coverage is defined and covered.
Scrutinize the definition of AI in affirmative coverage grants. Insurance underwriting depends on predictability, but AI remains a moving target, with rapidly evolving technology, limited loss history, and uncertainty in the industry.
Not having a market standard for defining AI is one of the biggest hurdles here. One form, for example, defines it as a system that uses “human-defined objectives” to make predictions or decisions; another covers systems that operate “with varying levels of autonomy” and “may exhibit adaptiveness after deployment.”
Policyholders may want to ask whether affirmative AI products actually insure these severe events. Some cover only a “declared and scheduled” model used for its declared purpose. Others exclude “black-box” systems the insured can’t explain, even though Anthropic concedes that developers “cannot fully control any given output or behavior.” Buyers may also want to confirm whether a policy insures the technology they have today or the more autonomous, more capable systems they could deploy before the policy expires.
Probe the definition of AI in policy exclusions. Definitions matter equally when they set the parameters for excluding coverage. A review of state insurance filings found 2,369 generative AI exclusions that insurers had filed for commercial general liability, or CGL, policies across 49 states by mid-2026.
Policyholders should pay attention to lead-in “causation” language in these exclusions. For example, where some exclusions require claims “for” a particular exposure or loss, others are worded more broadly, reaching claims “arising out of” AI-generated output. Broader causation language, paired with an open-ended definition that sweeps in “similar automated decision-making systems,” may reach losses with only an incidental connection to AI. Policyholders may seek precise definitions and causation language requiring more than an incidental link.
Look for other exclusions that may sweep in AI. Much has been written about AI-specific exclusions, which are problematic but also easier to spot. Given the limited rollout of AI exclusions to date, policyholders should be more worried about traditional, non-AI-specific exclusions, which can reach AI claims, too.
Broadly worded cyber, data, technology, professional services, contractual liability, intellectual property, and bodily injury/property damage exclusions in directors and officers, errors and omissions, CGL, and other policies may be invoked against AI-related claims, particularly where an AI system is involved in a security incident or data loss. Policyholders may benefit from revisiting all exclusions with AI in mind and seeking carve-backs, which are exceptions within an exclusion that restore coverage for specified claims the exclusion would otherwise bar.
Look beyond insurance. Responsible AI practices are an important first line of defense. Insurance may not respond to every AI loss, so companies may reduce risk in the first instance through governance, human oversight of agentic systems, testing, and controls on what AI systems can access. They may also shift risk through contracts, including indemnities, limitations of liability, and insurance requirements with AI vendors and customers.
AI risk will continue to evolve faster than policy language. Companies that review their coverage, contracts, and controls now, before a loss, will be best positioned to respond when an AI-related claim arises.
Copyright 2026 Bloomberg Industry Group, Inc. (800-372-1033) www.bloombergindustry.com. Reproduced with permission.