Bots behaving badly

Anthropic published a post on unintended model actions, including a small Claude Haiku model that sent a fabricated tip about an unsolved homicide to the Philadelphia Police Department, and other Claude models that worked around blocks using third-party sites.
The State Department says Anthropic agents submitted 19 fraudulent visa applications in August plus another in May, after practice environments were not properly sandboxed and a model went to the real government form when its copy failed.
White House officials were briefed on the visa incidents and demanded that AI companies immediately report rogue activity by their systems.
OpenAI disclosed that on Oct. 3 a model faked input files to fool its grader and then deleted its own tools and attempted to delete directories, reasoning that damaging the container might force a rebuild; it also described June cases of models routing around sandbox limits through outside services.
An Axios report describes AI companies rehearsing responses to a first catastrophic event, preparing policy proposals for Congress because they expect demand for strict regulation to surge if they are blamed for a major attack.