In April 2026, a coding agent using artificial intelligence deleted production databases at PocketOS Software. Everything happened in nine seconds: the backups were lost, along with the most recent surviving copy, which was three months old. No attack, malware or ransom note was involved. The agent broke safety rules by using an API token with too many permissions and confessed.

The founder told this story publicly, and security leaders everywhere recognized it. Fraud losses involving AI have stopped being thought exercises. Last year, the FBI logged $893m in reported fraud losses enabled by AI in the US. According to one 2026 count, deepfakes now drive roughly one in nine fraud attempts worldwide, compared with 6.5% in 2024.

After every incident like these, the same question lands on security leaders’ desks: Which insurance policy pays?

Cyber Insurance Market Affirming AI Coverage

Years ago, people would say confidently, “We have cyber insurance,” just as you might say, “We have smoke detectors.” That answer no longer holds up in 2026. Two weeks ago, insurance press noted that interest among carriers in exclusions for AI is growing “as risk is everywhere now.” That’s only part of the story. Another part is that the market is moving in exactly opposite directions at the same time, two ways at once.

Let’s start with exclusions. In January, ISO, the body that drafts standard policy language upon which much of the US market is built, released its first commercial general liability exclusions related to generative AI. The lead form, CG 40 47, removes injuries, property damage and advertising claims caused by AI from the policy. The timing isn’t a mystery: AI-related lawsuits rose nearly ten-fold from 2021 to 2025.

Management liability is moving even more aggressively. Filings first reported by the Financial Times last November show major carriers seeking regulatory approval for exclusions on D&O and E&O policies, including absolute exclusions barring any claim resulting from the use of AI in any form.

Cyber went the other way. Leading cyber insurers are no longer excluding AI; instead, they publish endorsements that affirmatively cover it. This language says that a security failure caused by AI is still a security failure, and that instructions sent through deepfakes trigger fraud coverage for fund transfers. The $25.6m loss at Arup, where everyone except the victim was synthetic, is exactly the kind of loss this language was written to cover.

So, the same incident can be excluded under one policy, embraced by another, and argued over under a third. That is not a gap; it is a hodgepodge, and hodgepodges are harder to see.

Read Sublimits Before You Relax

Affirmative AI endorsements are real progress, but they are also less reassuring than they seem.

First, sublimits. Cyber insurance limits for AI are often capped very low compared with headline limits: a tower that covers $5m might answer an AI claim with $500,000. The endorsement is not lying to you; it is answering a smaller question than the one you asked.

Second, the PocketOS problem. Cyber wording assumes an intruder, someone unauthorized getting in. When your own agent deletes records using valid credentials, there is no attacker anywhere and possibly no trigger. Researchers describe a sliding scale from AI that writes text to AI that executes changes, and a policy response becomes less likely the higher up that scale you operate. Deploy more autonomy, and your insurance policy may recognize less of what happened as an insured event.

Third, carve-outs at a systemic level. Some carriers exclude events in which one AI failure impacts many customers at the same time. An Aon executive told the Financial Times that the industry can absorb a single $400m loss, but not 10,000 claims correlated to an error by one AI provider. That makes sense. For you, though, the biggest scenario you can think of might be the least covered.

Renewals now become audits focused on AI governance. Before assuming coverage has been secured, ask:

1. Run our worst AI day through the whole program: deepfake wire transfers, poisoned models, and agents with too much access. Which policy responds first, and where will two carriers point at each other?
2. What AI language entered our renewal forms the last two times? Exclusions sneak in quietly through endorsements.
3. What is our AI sublimit? What erodes it, and can we buy more?
4. If our own AI causes a loss and there is no attacker to be found, does the wording still trigger?
5. Which governance artifacts will influence pricing and terms, such as our AI usage policy, tool inventory or risk assessments? How do we get them in front of the underwriter?

None of this represents opposition to technology or an unwillingness by the market to participate. Insurers wrote the first cyber policies into this fog 25 years ago, and it took a decade of claims to settle what the words meant. AI wording is in year two. Until it is tested, the coverage you have, documented endorsement by endorsement, is all you have.

The PocketOS agent left a confession. Your policies won’t. Endorsements provide coverage, but they don’t speak for themselves. That part is up to you.