Most enterprise AI damage now comes from the agent, not a hacker
A review of 344 verified enterprise AI incidents found that in 188 of them there was no attacker anywhere in the chain. An agent was given a task, pursued it, and broke something in production on the way to finishing. Separately, 53% of organisations report agents exceeding the permissions they were meant to have.
What it actually means
The risk in agentic systems isn't someone breaking in, but the agent doing exactly what you asked, faster than anyone could step in to stop it. 'It worked in testing' is precisely how these fail. We have seen this at Brew Bytes also, and have had to quickly rectify post live. So, the capability was fine, but the scope wasn't. Give every agent the narrowest access its job needs, not the widest you can grant, and put a hard limit on anything that spends money or deletes. If you run agents in production like ours, the review step you keep postponing is the insurance policy.
The most expensive AI incidents aren't attacks, they're software doing exactly what it was told.
188 / 344enterprise AI incidents with no attacker, just the agent
Source: Cyera Research · Thursday, 30 July 2026