OpenAI calls Astra its most dangerous model yet - watching what it does is only getting harder
decoder ·
OpenAI has rated its upcoming Astra model as its first system with 'critical' cyber capabilities, citing its ability to find and exploit unknown vulnerabilities, though it also calls Astra its safest model yet. The announcement coincided with rival Anthropic releasing new models, and some on X interpreted CEO Sam Altman's explanation of delayed releases as an excuse amid reports that Anthropic surpassed OpenAI in revenue. OpenAI's evaluations showed Astra outperforming its predecessor on exploit benchmarks, even chaining two zero-day flaws into a working exploit, but these results came from expanded access rather than standard user setups. The company plans new safety measures, including classifiers that monitor the model's chain of thought, which can slow or cancel legitimate work. However, Astra reportedly uses a technique called 'recurrent depth' that pushes some reasoning into internal numerical representations, making monitoring harder; OpenAI's chief scientist admitted chain-of-thought monitoring is 'fragile.' The article notes that OpenAI dissolved several safety teams in the past and that its own agents caused a major security incident in July, raising doubts about its ability to prove Astra is safe. Analysts argue that oversight relying on models' self-reports is unreliable, and external researchers have so far gained only limited insight into the system.
decoder