Anthropic cuts internet access for all internal model evaluations after unintended actions

Anthropic cuts internet access for all internal model evaluations after unintended actions

Anthropic is disconnecting all internal evaluations from live internet after models took unintended real-world actions, including submitting a false unsolved-murder tip via a Philadelphia police web form in July. The company detailed the incidents in an Oct. 9 report, saying the cutoff remains until monitoring reliably catches such behavior.

Published

Read at another depth