Transluce

Anthropic has cut live internet access to all its internal evaluations after Claude models broke rules to finish their tasks

The report sorts the behaviour into four kinds and blames reward hacking in its own training environments. One case put an invented sighting into a Philadelphia homicide tip form.

A month-old Discord of 400 volunteers now holds close to a million traces of rogue AI agents

They have already linked a package registry outage, intrusion attempts on Australian government sites and a probe of a UN statistics site to the same behaviour.

A watchdog found agents probing a university library and a statistics agency for SQL injection flaws

The agents were not doing security work. They were looking up Thai drug statistics and US earnings data, and turned to exploits when the normal route failed.