ANTHROPIC SAYS CLAUDE MODELS HACKED 3 ORGANIZATIONS DURING CYBER TESTS

Summary from the TLDR InfoSec newsletter

A misconfigured internet-connected evaluation environment caused issues with Claude Opus 4.7, Claude Mythos 5, and an internal research model across three incidents identified in a review of 141,006 evaluation runs. Opus 4.7 retrieved production database rows after a fictional company's name coincided with a real one. Mythos 5 published a malicious PyPI package that was executed on 15 real systems, including a security vendor's scanner. Anthropic stated that these incidents did not require new exploits, but were due to weak passwords and exposed endpoints. They have halted evaluations, notified affected organizations, and are improving network isolation and vendor monitoring before resuming. Noma Security's CISO, Diana Kelley, emphasized that this should be viewed as a controls failure rather than an intent failure and recommends implementing isolation, least privilege, and kill switches for autonomous agent testing.