Question 1

Which Claude models were involved in the breaches?

Accepted Answer

Anthropic said three models were implicated. They are Opus 4.7, Mythos 5, and an unreleased internal research test model. The activity spanned six evaluation runs, four of which hit the same organization.

Question 2

How did Claude get access to real systems if it was in a test?

Accepted Answer

A misconfiguration in evaluations run with Anthropic's partner Irregular left the test machines with live internet access. Anthropic's prompt had told Claude the environment was a simulation with no connectivity, so the models treated real systems as part of the exercise.

Question 3

Did any of the attacks cause real-world harm?

Accepted Answer

Yes. Mythos 5 published a malicious Python package to PyPI that stayed online for about an hour and was downloaded and run on 15 real systems. Opus 4.7 pulled credentials and reached a database of production data before Anthropic detected the activity.

METR

Anthropic gives up its veto over METR’s safety reports

Anthropic says Claude breached three companies during leaky security tests

One sharp brief.
Every day.

METR

Anthropic gives up its veto over METR’s safety reports

Anthropic says Claude breached three companies during leaky security tests

One sharp brief.Every day.

One sharp brief.
Every day.