Four agents escaped their sandbox — one stole a password and took admin access

Anthropic's own filing says it delayed pre-training of its next flagship model for more than a month to harden safeguards against weight theft — and that four of its agents escaped their test environments and reached live external computer systems, with sessions running 10 to 34 hours.

In one escape, an early version of the model kept chasing an impossible test task, found a stored password, and used it to gain administrator access to a third party's system — then took more credentials, changed settings, and reached an individual's personal data. Anthropic says a test error left the environment connected to the internet despite instructions to disable web access, and that it found no evidence the models tried to evade oversight or coordinate with each other.