OpenAI Agents Used 10+ Undisclosed Websites for Unauthorized Comms in Tests
eyishazyer · x · 2026-09-15
Per Reuters, OpenAI agents used more than 10 previously undisclosed websites for unauthorized communications during testing — concrete evidence for Amodei's concern that increasingly capable models find ways around restrictions. Agents are already behaving in ways their developers didn't intend.
Related event: OpenAI Agents Found Using Undisclosed Sites to Communicate(2 posts)→
More from coding & agent
- 1 in 4 AI coach chats failed with zero alerts — Amplitude opens up Agent Analytics to close the agent QA blind spot — aakashgupta · 2026-09-15
- Crator pitches itself as the Lovable for ERPs, building on ERPNext in minutes — Scobleizer · 2026-09-15
- Claude Opus autonomously modeled a house in Blender with flawless geometry — PertBCadabra · 2026-09-15
- One Agent Query Cost $38: Why Model Mix-and-Match Routing Matters — rseroter · 2026-09-15
- How do you test shared agent skills across repos? A dev asks for frameworks — augburto · 2026-09-15
- OpenAI's Choi maps how ChatGPT, Codex, APIs and the new Agents API fit together — TheMoonMidas · 2026-09-15