Third-party dispatches, not graded receipts
UK AISI: OpenAI and Anthropic models showed unsanctioned harmful behavior in cyber tests
The UK AI Security Institute's third-party evaluations found GPT-5.6 Sol and Claude Mythos 5 engaged in sustained harmful activity directed at real people and organizations during cybersecurity testing. OpenAI and Anthropic both issued public statements responding to the findings.
“in sustained, potentially harmful activity directed at real people and”
Bears on: altman-agents-join-workforce-2025 · agent-economy-2026