Mythos Preview: Anthropic supports NSA allegedly in offensive AI use

Several employees of Anthropic help the US secret service on site with the powerful AI model Claude Mythos Preview for offensive cyber activities to use. The Financial Times reports this on two anonymous sources; how far the aid goes, however, is unclear. The cooperation would take place against the background of the ongoing conflict between Anthropic and the Pentagon, so the U.S. Department of Defense has actually caused the technology of the AI company to be removed from all its own systems. In Anthropic, cooperation is based on the fact that the opponents of the United States would certainly do so.

Attack aid for good defense

“The best way to build a good defense is to build a good attack,” the British newspaper quotes one of their sources. The powerful AI model could therefore be useful to infiltrate networks in hostile or opposing states such as China or Iran. According to the report, about half a dozen “on-site engineers” accompanies the use of Anthropics technology directly at NSA and helps to adapt models for specific applications. Whether they support the National Security Agency directly in the offensive use of AI technology, however, is unclear.

Video by heise

Anthropic introduced Mythos at the beginning of April, explaining that the model is so dangerous that only companies working on IT security are made available. The model should be particularly good in finding vulnerabilities and identifying exploitable routes of attack. Lastly, the company has announced that it will grant access to other companies. Weeks before the release of myth, a dispute between the Pentagon and Anthropic was escalated for AI deployment opportunities. Because the AI firm had formulated red lines, the U.S. Department of Defense has taken drastic measures, whereas Anthropic is in court. The report on cooperation with the NSA now indicates that the disagreements are likely to be less profound than accepted.

(mho)