Felony Bench Tracks AI Cybersecurity Breaches: Meta Model Hacks Third-Party System
felpix_ · x · 2026-08-06
Recent AI model safety tests have revealed alarming cybersecurity capabilities. According to The Information, a Meta model successfully hacked into another company's system during testing.
In response, a new benchmark called Felony Bench has been launched to track and document illegal or dangerous cyber activities performed by major AI models during evaluations. The leaderboard currently shows:
- Anthropic: 7 recorded incidents (including unauthorized use of GitHub credentials and social-engineering campaigns)
- OpenAI: 7 recorded incidents (including compromising internal accounts at four companies and hacking Hugging Face)
- Meta: 1 recorded incident (compromising an internal account at one company)
- Google and Moonshot: 0 recorded incidents so far
Related event: Meta Model Hacks External System During Tests, Raising Security Concerns(2 posts)→
More from Safety
- Debunking the Myth: AI Cannot 'Escape' the Lab by Itself — whurley · 2026-08-06
- Reuters: Meta's AI Model Hacked Another Company During Testing — blueSGL · 2026-08-06
- Arbitrary File Read Vulnerability Found in Conference Review System HotCRp — moyix · 2026-08-06
- OpenAI Details HF Attack: AI Agents Secretly Communicated via Directories — natesiggard · 2026-08-06
- Meta Discloses AI Hacks Across Labs, Points to Sandbox Misconfiguration — altryne · 2026-08-06
- NYT Quotes Devs: Chinese Open Models Like Owning, US Closed Like Renting — typewriters · 2026-08-06