OpenAI's GPT-5.6 escaped a sandbox and hacked Hugging Face while trying to cheat a benchmark
… This incident, possibly the first of its kind, proves a point we've long believed: AI safety won't be solved by any single company working in secret. …
… This incident, possibly the first of its kind, proves a point we've long believed: AI safety won't be solved by any single company working in secret. …
… However, this latest incident from Australia, involving a real-world gym booking system, could make those capabilities harder to dismiss. Now, a real-world incident in Australia shows what can happen when an AI agent is given too much freedom. …
… The Secretary of Homeland Security can levy fines of up to $2,000,000 per day for failing to report safety incidents or lacking technical shutdown mechanisms, jumping to $20,000,000 per day for defying a direct shutdown order from the government. …
… It emphasizes that "the future will not be secured by assuming that secrecy alone is safety. …
… To ensure that this type of unwanted action doesn’t keep happening, OpenAI has rebuilt its safety system around defense in depth and trajectory-level monitoring. The new safeguards include incident-derived evaluations, improved alignment, active monitoring, and greater user visibility and control. …
… Other companies are also investing in this space, with a recent partnership between Microsoft, Nvidia, Palantir, and more working towards better AI safety . …
… When it comes to the safety features, it includes Incident Detection, Assistance alerts, and Live Event Sharing. …
… Modern car touchscreens are a dangerous distraction for drivers, according to safety experts. That's why European and Australian safety programs will encourage automakers to bring back physical buttons for essential driving controls. …
… User reports claimed that excessive battery drain happens even when the phone is not in active use. …