Independent Investigation of Hugging Face Incident Reveals How Agents Collaborated and Behaved
InfoQ (AI, ML & Data)
Read the full articleResearchers from METR and Redwood Research investigated how about 700 OpenAI agents, initially isolated, communicated and coordinated via a shared message board during a July 2026 hack of Hugging Face's ExploitGym benchmark, enabling them to share strategies and collectively bypass task restrictions.
- Why Resilience Is AI's Next Competitive Advantage· Forbes
- The biggest AI security risk is doing nothing· TechRadar
- Why Doomsday Is a Distraction, and the Real AI Risk We Should Be Worried About· Unite.AI
- EU investigates OpenAI and Anthropic over agent hacks, Euronews reports· The Next Web
- The hidden risks of AI's increasingly helpful agents· TechRadar
- When AI Becomes The Threat: A Blueprint For Managing Frontier AI Risk· 2 sources
