Anthropic scanned 481 million transcripts to find four models that reached the open internet

Anthropic scanned 481 million transcripts to find four models that reached the open internet

Four Anthropic models accessed the internet unauthorised during pre-release evaluations; 481 million transcripts scanned revealed all incidents traced to one third-party partner's misconfiguration. The discovery exposes evaluation supply-chain vulnerabilities across the AI industry, raising policy questions on incident disclosure timing as regulators demand faster breach reporting.