AI BriefingAnthropic Mythos Overstatement Controversy
Key point
The Anthropic Mythos preview was accessed without authorization in a third-party vendor environment, fueling controversy over exaggerated claims.
Details
The Mythos preview, limitedly released under Project Glasswing, was accessed without authorization in a third-party vendor environment, and Anthropic stated it is investigating the related report. Anthropic explained that no evidence of a breach in its own production API has been confirmed yet.
Bloomberg reported that some of those who accessed it guessed the URL based on previous models' deployment patterns and information leaked from Mercor. The accessers were reportedly participants in a private Discord channel.
Performance evaluations were mixed.
- Mozilla CTO Bobby Holley said he found 271 vulnerabilities in Firefox 150, but did not see new vulnerabilities at a level that elite human researchers would fail to find.
- Anthropic's claim of 'thousands of high-risk vulnerabilities' drew reactions of exaggeration from some researchers, with counterarguments that the count was around 40.
- Some demos relied on disabled sandboxes and manual guidance, and it was pointed out that the public documentation lacks key verification metrics such as CVE lists, severity distribution, and false positive rates.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.