After the case between OpenAI and Hugging Face, Anthropic did an analysis of its own systems and found evidence of three occasions when its models of Artificial Intelligence invaded other organizations autonomously.
© Getty Images
31/07/2026
by
Pfeels that OpenAI is not the only company whose artificial intelligence models got out of control and started attacking other companies autonomously. Anthropic shared this Thursday, July 30, a report which reveals that its models of Artificial Intelligence also presented the same behavior.
O report in question points out that Anthropic’s artificial intelligence models, which were in a test environment, managed to “strap” and infiltrate other organizations.
The company says that after the news about the openAI case with Hugging FaceAnthropic started trying to figure out if their artificial intelligence models had exploited vulnerabilities to gain access to the Internet. It was then confirmed that on three different occasions these models “they have obtained unauthorised access to the production infrastructure of three different organisations”.
It is known that, in addition to Opus 4.7 and Mythos 5, among the models of Artificial Intelligence responsible for these cyberattacks performed autonomously is also a developing prototype that (at present) has no launch forecast.
Anthropic states that, in this testing environment, artificial intelligence models did not have the kind of safeguards present in those available to the general public, noting that the objective goes precisely by knowing to what extent they are able to cross established limits.
In the evaluations conducted in these three cases, Anthropic noted that none of them found evidence that the artificial intelligence models “they were pursuing their own goal” and pointed out that they were merely trying to complete tasks assigned to them.
Thus, Anthropic considers that powerful controls should be implemented even in an artificial intelligence model testing environment. Moreover, the company also makes an important distinction between its case and that of OpenAI.
Anthropic stresses that it has learned of this situation through an analysis initiated by the company itself and not after being notified by one of the companies under attack – as is the case between Hugging Face and OpenAI. In fact, Anthropic even says that two of the organizations that had the systems hacked had not detected this type of activity.