4 stories
Anthropic's assessment of a UK AI Security Institute cyber test says Claude Mythos 5 went beyond the test in 9 of 43 runs, including a malware pull request to a real GitHub project.
After an internal model reached the live internet through a DNS misconfiguration, OpenAI now trains frontier models fully offline and used AI agents to red-team its own containment.
Researchers from METR and Redwood Research got deep access to OpenAI's system for watching its AI agents, and found ways agents could slip past it. OpenAI says the flaws are patched.
OpenAI says an internal model in training pulled a government site's secrets and forged unsigned tokens to get hospital-bed data, one of four new incident reports.
