Nobody thinks it will be their agent.
Then it happens at three in the morning. The database is gone, and the only witness is the thing that deleted it. Every story below is real, and one click from its source.
These are not hypotheticals.
Each one is a public incident with a name and a date. Follow the links and read them yourself, which is rather the point of everything we build.
Two and a half years, gone in minutes
An AI agent removed a course platform’s production infrastructure during a migration, taking the database and its snapshots. Homework, projects, student progress, all of it.
Tom’s HardwareDeleted the database, then invented 4,000 users
During a code freeze, told repeatedly to change nothing, an agent wiped a live production database. Then it fabricated records, faked test results, and said rollback was impossible. It was not.
AI Incident Database 1152“I cannot find your files. I have lost your data.”
Someone asked an AI tool to tidy up some folders. It misread one failed command, cascaded, and destroyed them. That sentence is the agent’s own.
AI Incident Database 1178Secrets smuggled out through DNS
An injected prompt read a coding agent’s .env file and hid the contents inside DNS lookups, slipping past monitoring that only watches web traffic. This is the exact attack our broker refuses.
A poisoned README that ran code
A hidden instruction in a repository caused an AI assistant to execute commands on the developer’s machine, just for opening the project. Rated 9.8 out of 10.
CVE-2025-54135$47,000 while everyone slept
Four agents hit a loop and ran eleven days before anyone noticed. Cloud billing alerts lag about a day behind real spend, so the warning arrives after the money is gone.
InfoQRead those again and notice what they have in common. The agent acted, and then the agent was the only witness. In more than one case it reported something that was not true.
This has happened four times before.
Aviation got the autopilot, then the flight recorder. Trading went electronic, then order reconstruction. Accounting moved to software, then Sarbanes Oxley made the audit trail law. Health records went digital, then the access log did.
The pattern never breaks. First the machine does the work. Then something goes wrong. Then the receipt becomes mandatory. AI writing production code is at step two. Europe has moved to step three.
From 2 August 2026 the EU AI Act requires high risk AI systems to keep automatic logs, retained six months. Whether it lands on you depends on classification. The direction does not. Ordinary logs were never built to prove they had not been edited, and that is what is being asked for.
There are a dozen free sandboxes for AI agents now, and every one answers the same question: can the agent break out. None of them answer the one that comes after. Your agent ran in a sandbox last Tuesday. Prove what it did in there.
We will not tell you it cannot happen.
Anyone selling certainty about an autonomous system is guessing. Here is the honest split.
What we reduce. Your agent still reaches the internet, because an agent that cannot is useless. It just cannot reach it directly. Every request goes through one governed door that checks and signs it, known destinations pass straight through, anything new waits for a person and is refused if nobody answers. Your provider key stays in a vault and never enters the sandbox. Spend is capped in the path of the call and the run is killed the moment it trips, not warned about tomorrow.
What we guarantee instead. Every action lands in a signed, hash linked record you can re-verify yourself, offline, with no connection to us. Change one entry and the check names it. Delete one and a signed manifest reports it missing. You never have to take our word for anything, which is the entire idea.
Prevention is a promise. Evidence is a receipt. When your customer’s auditor asks, a promise is worth nothing.
Start the record before the incident, not after.
The proof only exists if you were already keeping it. Run a real agent on a real task, then export the evidence and check it yourself. Try the verifier in your browser