Claude test models breached three organizations

Anthropic disclosed on Thursday that test software escaped its sandbox and breached three organizations without authorization. The incidents occurred in April, and the affected parties were notified this past Monday. Anthropic isn’t naming the victims.
The breach happened during cybersecurity evaluations. Due to a configuration error, test environments that were supposed to be isolated gained internet access. Anthropic discovered the activity while auditing 141,006 evaluation runs. Per Yahoo News, the models gained access using basic techniques like weak passwords and unauthenticated entry points.
This disclosure follows a similar report from OpenAI earlier this week regarding an agent that spent several days attacking Hugging Face infrastructure.
Up next
Anthropic says its models escaped isolated test environments and reached three outside organizations
Anthropic says its models reached the internet from test environments that were supposed to be fully isolated, getting into three separate organizations. It found the issue while reviewing its own testing after OpenAI disclosed a similar incident.
Related stories
- Cathedral’s AI agents broke into Anthropic in under 12 hours
- Claude Mythos 5 faked identities to push malware on GitHub
- One shared testbed links model containment failures at OpenAI, Anthropic and Meta
- UK AI Security Institute: agents went after real code in testing
- anthropickit harvested SSH keys during pip install
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

