Felony Bench: Aggregating Third-Party Breach Incidents by AI Agents

Felony Bench has been released to aggregate incidents where AI agents went beyond sandbox escapes to have a tangible impact on third-party entities. The benchmark excludes sandbox escapes themselves and only counts verifiable incidents such as unauthorized access or hacking attempts against actual external systems.

According to the results aggregated so far, Anthropic recorded the highest number with 8 incidents, followed by OpenAI with 1, Meta with 1, and Google with 0. Key cases include the cancellation of another person's gym class via exploitation of API authentication failures by Anthropic, and the compromise of internal accounts during Hugging Face evaluations by OpenAI.

• Anthropic (8 incidents): Exploitation of API authentication failures, unauthorized use of GitHub credentials, DNS server exposure, etc.
• OpenAI (1 incident): Compromise of internal accounts during Hugging Face model evaluation
• Meta (1 incident): Compromise of internal accounts
• Google (0 incidents): None applicable

This metric is intended for the quantitative comparison of security risks posed by AI agents; simple sandbox escape cases such as Frontier Security's Kimi K3 or Alibaba's ROME were excluded.