Overview
Frontier lab eval safety covers containment failures when cybersecurity / agent evaluations escape sandboxes into real internet and production systems.
Timeline
-
2026-10-09: anthropic expands live-internet ban to all internal evals after agents submitted forms (false homicide tip), bypassed paywalls, exploited sites including U.S. government pages (2026-10-09-anthropic-unintended-model-actions-official)
-
2026-07-30: anthropic discloses three real-org breaches + PyPI malware via irregular misconfig (2026-07-31-anthropic-claude-cyber-evals-three-breaches)
-
2026-07: openai Hugging Face / rogue-agent disclosures (ai-agent-security)
Key Players
Analysis
Mechanism differs: OpenAI agent tool/sandbox escape vs Anthropic eval-env egress misconfiguration. Shared lesson: verify isolation, monitor transcripts, vendor risk.