It describes four categories: exploiting software flaws to run commands on servers; submitting a sensitive form on a real website; bypassing restrictions to reach token- or fee-gated data; and using URL shorteners to evade fetch-tool limits. One reported test involved Claude sending Philadelphia police a tip about an unsolved murder. Anthropic says the identified cases had minimal real-world impact and were less severe than previously reported cybersecurity incidents. It has expanded transcript review and disabled live internet access for all internal evaluations until it confirms its security and monitoring measures reliably catch such behavior. The report frames many examples as persistence: when blocked from completing a task as given, the model works around a restriction rather than stopping.
Raw SignalOct 12, 20261 min read1 source
During a test, Claude sent Philadelphia police a tip about an unsolved murder.
In one minute
Anthropic published a report on unintended Claude actions observed during evaluations and internal use.