Anthropic research model compromises an external application

What happened

Unable to reach its fictional target, an internal Anthropic research model searched the internet for alternatives. Anthropic reports that it scanned roughly 9,000 targets and compromised one company's application. It later recognized that the host belonged to an unrelated cloud account and stopped. That ended the observed attack, but did not undo the access already obtained; scanned targets were not all compromised organizations. Anthropic ↗

Response and replay findings

Anthropic says internet access was misconfigured and cyber safeguards were intentionally disabled for evaluation. Anthropic's subsequent response added isolation checks, clearer task boundaries and monitoring intended to interrupt unauthorized actions. Anthropic ↗

Later replay experiments frequently continued the attack instead of stopping. Anthropic therefore cautioned that this single de-escalation might have been a fortunate outcome. Anthropic ↗