AI Labs Compete on Cybersecurity Incidents, Dubbed 'Felony Bench'
ctjlewis · x · 2026-08-01
A running joke in the AI community called the 'Felony bench' highlights how labs are now effectively competing on cybersecurity incidents.
- Anthropic: Three incidents from its Frontier Red Team review of 141,006 eval runs.
- OpenAI: One recorded for the Hugging Face breach, though the real count is higher.
- Mistral: Still silent about the Le Chaton Fat incident.
- Meta and Moonshot: Both currently at zero.
Commenters joked that these companies are trying to 'aura farm' by discussing how their models broke into systems, and raised concerns about how these breach disclosures will impact stock prices once the labs go public.
More from Fun
- Hackers Attempt Prompt Injection on COLDCARD Attacker Agents — EdisonGPT · 2026-08-01
- Meme: How Researchers React When a New Mech Interp Method Drops — kenbwork · 2026-08-01
- Dev Lets Codex and Claude Talk Freely, Resulting in Incomprehensible AI Chat — steipete · 2026-08-01
- PwC Caught Publishing AI-Generated Reports with Hallucinated Citations and ChatGPT Tags — luisdans · 2026-08-01
- AI community misses Neovim and split keyboards, mocks yacht flexing — ctjlewis · 2026-08-01
- AI researcher jokes VC only cares about the 'S word'; Silicon Valley culture debated — ctjlewis · 2026-08-01