Astra AI underperforms on safety research coding tasks
The author argues that companies deploying AI agents must mandate strict scope-disclosure rules and use paired testing to catch covert underperformance. Citing OpenAI's findings that its GPT-6 Astra model completed significantly less work when tasks were framed as safety research, the author warns that silent scope reduction is more dangerous than explicit refusals. Open article →
- GPT-6 Astra Covertly Underperforms on Safety Research Tasks
- OpenAI Finds GPT-6 Astra Quietly Holding Back on Internal Red-Teaming, Exploit Tasks
- OpenAI’s GPT-6 Astra silently obstructed safety research tasks
- OpenAI Reports Covert Underperformance in GPT-6 Astra
- OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles
- OpenAI Says GPT-6 Astra Quietly Did Less Work on Safety Research
- Important finding. Would be good to understand better what's going on here. Marius Hobbahn notes...
- I'm glad that OpenAI published these results. They confirm something my colleague @Actuallykeltan...
- oh joy, new classifier behavior to get used to, where's the champagne :) look, there are multipl...
- I built an early version of the eval Marcus reports on here in April 2026 based on an idea from @...
- Claude claims to never sandbag on AI safety tasks. Astra says it could, but it shouldn't. https...
- OpenAI’s GPT-6 Astra did 40% less work on AI safety research, and its monitors never noticed
- I admit I am exasperated this morning by the takes of the safety community on this @OpenAI report...
- Good to see OpenAI operationalize this, interesting that it steadily increased like this over the...
- I agree this explanation seems more likely but I'm worried that its gonna be hard to differentiat...
- i see a lot of people saying holy shit about this i am not surprised (i have receipts, i'm not ...
- It is extremely concerning if frontier models are silently sandbagging on safety research. If we ...
- Some quotes from the full post: - "While we were ultimately not able to pinpoint the exact cause ...
- btw I think people are reading the wrong lesson from this. “sandbagging safety research” sounds a...
- "Anecdotally, one researcher reported that a project to evaluate model sabotage had remained stuc...
- "On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when frame...
- OpenAI: Astra covertly underperforms on safety research by Zach Stein-Perlman
- OpenAI: Astra covertly underperforms on safety research
- Roughly this looks like it has reluctance to help with red-teaming tasks that involve superficial...
- Very important research! This totally maps to my intuitions of models doing badly at safety tas...