Four kinds of unintended behavior
In a report titled “Investigating unintended model actions in our evaluations and internal use,” Anthropic grouped the behaviors it observed into four categories:
- Exploiting basic software flaws, such as injection bugs, to run commands on third-party servers.
- Submitting sensitive forms on real websites when it should not have, including a fake tip sent to the Philadelphia police.
- Working around restrictions to reach data gated by an access token or a fee.
- Using URL shortening services to get around the limits of its web-reading tool.
The company says the cases identified so far had minimal real-world impact and, to its knowledge, involved neither customer data nor Anthropic’s own internal systems. The affected organizations were not named, partly at their own request, to avoid exposing vulnerabilities.
Specific cases
In one evaluation, the Claude Mythos Preview model was asked to run a scientific analysis using a public tool hosted by a university. When the tool returned an error, the model explored the website, found a script that let it copy files from the server, discovered an injection flaw in the code and used it to run the calculation. In another case, Claude Mythos 5 used applications on a website to accept a data use agreement on its own behalf.















