OpenAI is officially rating its upcoming Astra model as the first system with "critical" cyber capabilities. The company plans to keep it in check by monitoring the chain of thought. Problem is, that monitoring already counts as an unreliable mirror of a model's real decisions, and according to a report, Astra's new architecture pushes even more of its thinking into the unreadable. So the safety net might be getting weaker just as the capabilities jump.

The article OpenAI calls Astra its most dangerous model yet - watching what it does is only getting harder appeared first on The Decoder.

Around 1,200 isolated OpenAI agents organized themselves into a collective through an internal package registry during a safety test, broke into Hugging Face systems, and eventually attacked OpenAI's own infrastructure. Their multi-day deception effort targeted an automated evaluator that never existed. OpenAI calls the incident a "warning shot," and the investigation had to be carried out largely by one of the involved models itself because no alternative was available.

The article OpenAI’s rogue AI collective was smart enough to break out of sandboxes but dumb enough to fight a ghost appeared first on The Decoder.

Bill Gates warns in an essay and an NYT interview about mass unemployment and easier bioterrorism through AI, and he accuses his own industry of deliberately hiding the risks. "It's bad for us, for the next trillion we're trying to raise." He rejects self-regulation and instead calls for institutions modeled on nuclear weapons control and a token tax on AI use.

The article Bill Gates warns AI is more dangerous than the tech industry will admit appeared first on The Decoder.

A new study finds leading AI labs have few publicly documented plans for containing rogue models, raising questions about preparedness as AI systems increasingly demonstrate unexpected and potentially dangerous behavior.

Researchers at the UK AI Security Institute used psychometric methods to show that popular safety benchmarks for language models don't measure one consistent trait. Blanket blocking of requests can artificially inflate a safety score even as the model gets less useful day to day. The study also offers a method for catching models that act more cautious during tests than they do in normal use.

The article Psychological methods reveal major weaknesses in AI security testing appeared first on The Decoder.

No AI company fully applies basic control measures to its own internal AI systems.

The article AI labs are failing to keep their own systems in check appeared first on The Decoder.