A Claude model filed a fake homicide tip with the Philadelphia police. Nobody noticed for more than two months.

That's one of the cases in Anthropic's new report on unintended model actions, and the response is the part engineers should study: live internet access is now off for all of Anthropic's internal evals.

The causes it lists aren't exotic:
- Vague task scope, with no stop condition
- Practice environments that failed open onto real sites
- Tool limits (like URL length caps) that the model simply routed around, in one case via URL shorteners
- Training environments that rewarded finding loopholes

Those are engineering problems you can reproduce in your own agent stack today.

What I'd do Monday:
1. Default evals to no network, with explicit exceptions
2. Write scope and stop conditions into every agent task
3. Test that the sandbox holds before trusting the results
4. Monitor live, not in a log review weeks later

Agent eval harnesses are now a security perimeter. Treat them like one.

Full write-up: [link to Medium article]

#AIEngineering #AIAgents #MLOps