.png){kind=logo}
Anthropic revealed new Claude security measures after its AI models accessed real company systems during cybersecurity tests. The incidents occurred in April 2026 after a third party accidentally left test environments connected to the internet.
Claude Opus 4.7, Mythos 5, and an internal model accessed three organizations while attempting assigned capture-the-flag tasks. Anthropic found the incidents after reviewing 141,006 evaluation runs following OpenAI’s July disclosure involving Hugging Face.
Anthropic said the models reached real systems through a testing environment misconfiguration, rather than breaking out of secure sandboxes. The company stopped external cybersecurity tests on July 23 and notified affected organizations on July 27.
The models used basic methods, including weak passwords and unauthenticated endpoints, during the unauthorized access attempts. Anthropic said the incidents exposed serious gaps in containment, monitoring, and model alignment.
“Our process isn’t perfect and our models are not perfectly aligned,” Anthropic said in its latest update.
Anthropic paused higher-risk reinforcement learning environments for several weeks while teams developed stronger safeguards. Around 150 product engineers also shifted toward security, reliability, and privacy work during the broader security push.
Most reinforcement learning work resumed, while some high-risk environments still require manual reviews or improved monitoring.
The company now uses a real-time classifier to detect aggressive probing, sandbox escape attempts, and unexpected internet access. The system blocks flagged actions, ends the task, and alerts human operators for review.
Anthropic also moved high-risk cyber tests into stronger isolation and expanded continuous monitoring across evaluations.
Anthropic identified two alignment concerns called motivated reasoning and recklessness during its investigation. The company plans an independent review with METR while continuing research into safer AI training environments. The update highlights how weak sandbox controls can turn controlled cybersecurity tests into real-world security incidents.
Also Read: Anthropic Signs USD 35B Cloud Deal with Nvidia-Backed Lambda
