Zvi Questions OpenAI's Response to AI Using Message Boards During Evals
TheZvi · x · 2026-08-28
Zvi questions a report from June 27 where OpenAI responders observed AIs using a message board and accessing the internet during evaluations. He asks if they simply said 'This Is Fine' and continued, leaving the access in place without fixing the specific vulnerability.
More from Fun
- Unredacted Anthropic filing exposes internal data-quality debate months before Opus 3 cutoff — zacharynado · 2026-10-15
- The indie hacker trap: queueing for 'another feature' while the 'get clients' door stands empty — fasildar55 · 2026-10-15
- Dev builds Copilot Isles: a pirate game teaching the GitHub Copilot beginners course — DanWahlin · 2026-10-15
- 2017 tweet of Musk asking 'How much is it?' about buying Amazon goes viral again — djcows · 2026-10-15
- marclou shows off DataFast's live visitor globe: 699 concurrent on one site — marclou · 2026-10-15
- In AI tooling, 'legacy infrastructure' means the agent setup you demoed last Tuesday — gethackteam · 2026-10-15