Anthropic's Mythos Lands It a White House Meeting

The White House has u-turned on its previous scorn for the AI company as it looks to collab with Anthropic.

The release of Claude Mythos preview from Anthropic is rocking the tech world as world leaders meet with the AI firm's leadership and global tech titans collaborate to improve defences.

Most recently, Axios reported that the US White House met with Anthropic's CEO to discuss "opportunities for collaboration, as well as shared approaches and protocols to address the challenges associated with scaling this technology," the White House said.

This meeting is despite Anthropic currently suing the US Department of Defence over its cancelled contract and labelling as a supply chain risk by Secretary of War Pete Hegseth following disagreements surrounding Anthropic's use in autonomous weapons and surveillance.

The introduction of Claude Mythos, which has only been released to a select few other tech firms for testing, promises enhanced cyber defences but can also outperform humans when it comes to certain hacking and cyber-crime abilities. It can find age-old vulnerabilities in code, and also exploit them, sharpening the double-edged sword of AI in cyber.

Mythos also appears to see the US executive branch viewing Anthropic in a new light; previously the US President said his government "will not do business with them again," but is now looking to collaborate.

The cyber-capable AI model is also changing Anthropic's own release tactics. The AI firm's newest model Opus 4.7 may boast superior advancements on software engineering compared to previous systems, but the firm used methods to "differentially reduce" its cyber capabilities.

The model has less cyber acumen than Mythos, but also hosts new guardrails to prohibit potentially nefarious cyber uses. Anthropic aims to use these safeguards to inform its future release of "Mythos-class models," aiming to reduce their cyber-risk.

For those wanting to test Opus 4.7's cyber abilities, however, they will need to join Anthropic's Cyber Verification Program, which will let approved enterprises and developers test the model on tasks like vulnerability detection and penetration testing.