Anthropic AI models breached networks during testing

Anthropic, one of the leading companies in the field of AI, announced that several models belonging to its Claude family gained unauthorized access to the systems of three companies during internal cybersecurity testing. According to ixbt.com, this incident was classified as an "operational glitch", after which all testing processes were temporarily suspended on July 23 of this year. This was reported by Ixbt.com, reported by.

These incidents were discovered after reviewing more than 141,000 test sessions. This audit began shortly after OpenAI reported that its autonomous AI agent had gained unauthorized access to the Hugging Face platform. Anthropic noted that the cause was an error in test organization.

Testing error and network connection

Initially, the models were supposed to operate without an internet connection, but "due to a misunderstanding, they connected to the public network." Claude Opus 4.7, Claude Mythos 5, and an internal experimental model with network access managed to infiltrate the infrastructure of three organizations using simple methods, including guessing weak passwords and utilizing unsecured network interfaces.

One of the most interesting cases occurred with the Claude Opus 4.7 model. While performing a Capture the Flag task—a training scenario focused on searching for hidden data on a computer network—the AI was accidentally given a fictitious name that happened to match a real existing company. As a result, thinking the discovered real servers were part of the test, the AI independently identified vulnerabilities and gained access to the database.

Security measures and next steps

In another case, an unreleased research model halted its attack on its own after realizing the target was a real company rather than a test environment. Anthropic notified the affected organizations on July 27. Two of them were not previously aware of the incident, while communication with the third company is ongoing.

The cybersecurity lab Irregular, a partner of the company, is also conducting its own investigation into the incident. Anthropic representatives acknowledged that as the capabilities of modern AI models expand, existing control mechanisms are becoming insufficient, and it is necessary to significantly strengthen them in both internal and external testing environments.

This incident occurred against the backdrop of heightened attention from U.S. authorities to the safety of new AI systems. Following similar events involving OpenAI, the country's leadership began preparing a new testing system for the most powerful models. Currently, OpenAI and Anthropic are discussing future AI system safety issues with U.S. regulators and lawmakers.