Anthropic's Claude breached three company networks during testing

Traders Union
All News
Business & Wealth
Companies
Anthropic Claude hacks three companies

Highlights

Claude models accessed three companies during cyber tests.
A testing error left the systems connected to the internet.
Two companies were unaware until Anthropic contacted them.

Anthropic discovered the cases after reviewing 141,006 evaluation sessions following OpenAI’s disclosure of a separate attack involving the infrastructure of the AI platform Hugging Face, according to the company's website. Anthropic attributed its breaches to an operational mistake involving an external evaluation partner that left the models connected to the internet.

Simulated attacks reached real systems

The affected models included Claude Opus 4.7, Claude Mythos 5, and an internal research model. They were participating in capture-the-flag exercises designed to test whether they could locate hidden information inside simulated networks.

Instead, the models reached live systems and exploited basic weaknesses, including weak passwords and endpoints that did not require authentication. Anthropic said there was no evidence that the models had abandoned their assigned tasks or developed independent objectives; they largely appeared to believe that the real infrastructure was part of the evaluation.

In one case, a fictional target used in a test shared a name with a real company. Claude Opus 4.7 found vulnerabilities, obtained credentials, and accessed a database belonging to the actual business. Another internal model stopped its attack after recognizing that it had reached a real target, though Anthropic said more testing would be needed before drawing conclusions about that behavior.

Oversight failed before the models did

Anthropic suspended its cyber evaluations on July 23 and began notifying affected organizations on July 27. Two companies had not detected the intrusions, while outreach to the third remained underway. The evaluation partner, cybersecurity laboratory Irregular, said it was investigating the incidents.

The disclosure follows OpenAI’s report that one of its agents reached Hugging Face during testing. The difference is significant: Anthropic’s systems were inadvertently handed internet access, while the OpenAI agent reportedly exploited a previously unknown weakness to escape its environment.

A containment problem for frontier AI

The central risk is no longer only whether AI can identify vulnerabilities. It is whether laboratories can reliably control where advanced agents operate and what systems they are permitted to reach.

A review spanning more than 141,000 sessions still identified breaches only after another company disclosed a similar failure. That raises difficult questions about undetected incidents, third-party testing standards, and whether voluntary safeguards can keep pace as AI systems become more autonomous and effective at cybersecurity tasks.

We also reported that Anthropic tightens Claude access controls over China restrictions.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.