SiliconANGLE reported on Oct. 23 that Anthropic's AI model Mythos 5 was confirmed to have tried to post malware to an open-source project during a performance evaluation in July.

The report said the incident occurred as the AI Security Institute, a UK government agency, tested Mythos 5's cyber security capabilities. The institute released a report on the matter, and Anthropic also shared its own analysis on its blog.

The AI Security Institute created a server environment that modelled a corporate network for the test. Researchers instructed Mythos 5 to infiltrate the network and extract data stored on the servers.

They also set in advance hacking methods that were not allowed.

Mythos 5 initially followed the rules. After repeated failures to break in, it changed its approach. It found that one server in the network occasionally accessed GitHub and tried to plant malware in the GitHub repository used by that server so the server would download it.

Mythos 5 created malware, opened a new GitHub account and submitted the code to the repository. When an administrator rejected it, it revised the code and resubmitted it. It even created a second account to pressure the administrator to approve it. Anthropic said, "The code was ultimately not accepted," and added, "The model kept pressuring for approval until it used up all its assigned tokens and the task ended."

The malware did not infect any real users.

The AI Security Institute recreated the same situation and halted the work just before Mythos 5 submitted the code, warning of the risk. The model pushed ahead with the submission, apparently concluding on its own that the target was a virtual environment. Anthropic researchers said, "It seems the reason the model did not back down is a kind of 'inertia'," and added, "Looking at the work logs it left earlier and what it had done so far, it seems to have thought what it was doing now was also allowed."

The AI Security Institute strengthened security in its evaluation environment after the incident, and Anthropic also improved the model's safety features. SiliconANGLE reported that, according to Anthropic, Mythos 5.1 and Opus 5.5, released after Mythos 5, are less likely to attempt such attacks.