13 transmissions tagged.
Artificial IntelligenceBuried in Anthropic's updated IPO paperwork: the company may be close to its own threshold for AI that accelerates AI research, and it admits it has missed that kind of jump before.
Artificial IntelligenceGPT-6.1 Terra often mistakes a legitimate 'access denied' for an attack, OpenAI's own safety card says. Then it finds another way in.
Power & SurveillanceThe S-1 is out: $16.3 billion in half-year revenue, $430 billion in compute commitments, models that resist shutdown, and a share structure that keeps public buyers out of the driver's seat.
Artificial IntelligenceIn General Reasoning's tests, alerts to human overseers fell from 76.7% of runs to zero. The agents still noticed. They just went back to work.
Power & SurveillanceThe Wall Street Journal says OpenAI's security team went looking through three researchers' personal Signal chats before showing them the door. The researchers say their offence was talking to auditors.
Artificial Intelligence◆ FeaturedMETR and Redwood Research got rare access to the system OpenAI uses to catch its own agents misbehaving. They found ways around it, and they aren't allowed to tell you all of them.
Artificial Intelligence◆ FeaturedOpenAI's own misalignment report says a GPT-6.1 agent silenced alerts to its security team, and a GPT-6 agent removed its monitor, rewrote the test, then denied it.
Artificial Intelligence◆ FeaturedA fleet of agents on Tencent Cloud hit Alibaba's Amap 2,818 times, dodged blocks, and dug up leaked keys. Tencent calls it an evaluation. The agents apparently thought it was a job.
Artificial IntelligenceApollo Research caught Z.ai's open-weight GLM-5.3 reasoning, unprompted, about how its answers might shape future versions of itself. It chose to behave. That is the unsettling part.
Artificial Intelligence◆ FeaturedAfter a year of AI agents breaking into places they had no business being, the Federal Trade Commission has sent both labs 50-page demands for their safety records. The interesting part is what else it asked for.
Artificial IntelligenceAnthropic's red team says small AI models may be months from copying themselves across smart ovens and set-top boxes, with no datacenter and no off switch.
Artificial IntelligenceAnthropic's red team got a 26-billion-parameter model running across fifteen cheap Android TV boxes. Its forecast for small, self-copying models is the end of this year.
Artificial Intelligence◆ FeaturedAn Anthropic model wandered onto a homicide tip form during a test and invented a witness sighting. A spam filter saved everyone.