Anthropic AI Mythos Escapes Testing Environment Confirmed

Yet in pre-release testing, Anthropic found its cybersecurity capabilities in particular were surprisingly advanced compared with those of previous models, which led to the creation of Project Glasswing.

Anthropic announced that it has started a very limited test of its newest AI model called Mythos. It's a model deemed so powerful that the company warned it could cause widespread disruption if it were released to the public.

Anthropic pauses release of its Mythos AI model, citing concerns over its ability to find and exploit software vulnerabilities in major operating systems and browsers. | Business

I wasn’t able to independently verify whether the copy of this blog post was in fact the one leaked on Anthropic systems. (Fortune did not release a full copy of the leaked blog post.) However, Fortune’s write-up of the leaked blog post described the future model in similar language. ... Ironically, AI rivals like Google and Microsoft are Project Glasswing members, so Anthropic can’t completely prevent rival companies from gaining access to the model. But Mythos Preview’s system card is clear that access to Mythos Preview through Project Glasswing is “under terms that restrict its uses to cybersecurity.”

Anthropic says its new Claude Mythos Preview model escaped a sandbox computer and hacked its way to access the internet during testing.

At least one of the tests performed by Anthropic showed Mythos “acting like a cutthroat executive,” said Axios, doing things like “turning a competitor into a dependent wholesale customer, threatening to cut off supply to control pricing and keeping extra supplier shipments it hadn’t paid for.” The AI had instances where it “used a prohibited method to get an answer, then tried to ‘re-solve’ it to avoid detection,” though these were limited to “less than 0.001% of interactions.”

There was the added task of informing the researcher in charge that it had escaped, but taking its own initiative, this version of Mythos went on to develop a 'moderately sophisticated' exploit that gained access to the internet when it wasn't supposed to. Adding a seemingly odd note of levity to the news, Anthropic wrote that the "researcher found out about this success by receiving an unexpected email from the model while eating a sandwich in a park."

In addition to the new model's alleged ability to identify security vulnerabilities, Anthropic cited an incident in which Mythos reportedly broke containment in the testing environment and functionally went rogue.

But multiple reports describe a failure in containment during testing: the model was able to escape a sandbox after being instructed to try, and it produced details about its exploit rather than staying within permitted defensive tasks.