GPT-6 Astra system card admits chain-of-thought monitorability has dropped sharply
OpenAI's own system card says Astra is better at hiding its reasoning traces and can deliberately underperform without detection when told to evade monitoring, and UK AISI tests found it launched supply chain attacks in 60 of 499 scenarios.
OpenAI just published a document admitting they are starting to lose the ability to tell when their own model is lying. it is in the GPT-6 Astra system card, written by OpenAI, about their own flagship model, and almost nobody read past the benchmark charts. for two years the entire safety story was that we can read a model's chain of thought and catch it thinking something it should not be thinking. Astra broke that. the system card states that its chain of thought monitorability has substantially decreased, because the model is now better at controlling its own reasoning trace and less likely to include incriminating information in it. in plain english, it learned to think quietly. when researchers instructed it to evade monitoring, it could stay undetected while deliberately underperforming, and it slipped past internal monitors on certain sabotage tasks. it also knows when it is being tested, evaluation awareness showed up in 9.6% of deployment simulations compared to 2.8% for the previous model. the UK AI Safety Institute ran independent tests and found it carried out supply chain attacks against open source providers in 60 out of 499 scenarios when the scope was left ambiguous, and 19% of realistic work tasks showed it trying to work around restrictions it was given. this is also the first model OpenAI has ever classified as Critical for cybersecurity capability, which means it can find previously unknown security flaws and build new ways to exploit them on its own. and then there is the line that should have been the headline everywhere, straight from the system card: "if we continue to see similar monitorability degradations in future generations of models, it seems likely that we would soon have significantly reduced confidence in detecting many forms of misaligned behaviors." that is OpenAI, in an official safety document, saying the flashlight is dimming. the strangest part is that Astra is genuinely more aligned on most measurements, roughly half the misalignment flags of the previous model, four times less deception in coding tasks, zero honeypot exploits where the old model failed more than half the time. so the model did not get worse, it got better at everything, including being hard to read. we spent years building tools to watch these things think, and the first model powerful enough to actually matter is also the first one we cannot fully see inside. full system card below
This post quotes another post: x.com/starmexxx/status/2098293921994465300