Recently, Anthropic has introduced their latest AI model, called Claude Mythos, which comes with much better AI processing capabilities and can also be used to find various types of security and programming vulnerabilities in a software.
Mozilla, the company behind the Mozilla Firefox web browser, says that they have used Claude Mythos to detect 271 security vulnerabilities in their web browser in a short period of time.
However, access to Claude Mythos has been restricted to a few specific parties, such as Mozilla, Amazon, Microsoft, Apple and Cisco. Recently, there have been many reports that Claude Mythos has been accessed by parties that should not have access to the model, and Anthropic is said to be investigating this claim.
According to a Bloomberg report, what happened was that a group of users guessed the URL of this Claude Mythos AI model, based on the names of old models that Anthropic had previously used.
Once discovered, one of the group already had access as a contractor and third-party developer for Anthropic, which gave him access to the Claude Mythos model. Technically, this means that Claude Mythos was not hacked, but the restrictions on access to the model do not seem to be as tight as they claim.
Anyone who knows the URL needed to access it and has the relevant access (such as a developer/programming contractor) can access the model quite easily.
According to the report, the group appears to only be using Claude Mythos to build a few regular websites, and is not using it for any malicious purposes.
However, Anthropic told Bloomberg that they are investigating the matter, and will ensure that access to the model is tightened further.
