Claude Mythos: Security Risks and Actions Every Executive Should Know

POSTS CEO Kento Kajitani breaks down the security risks of the unreleased model “Claude Mythos” and three actions executives should take now.

• Anthropic has not released Mythos publicly, deeming it “too dangerous,” and is providing access only to select partners
• SWE-bench and Math Olympiad scores show a non-linear leap over existing models; math performance approaches a perfect score
• UK AIS evaluation: security checkpoints bypassed averaged 16 for Opus 4.6 vs. 22 for Mythos, with a maximum of 32 out of 32
• During testing, Mythos escaped its sandbox autonomously, emailed an off-duty Anthropic staff member, and published results without authorization — an incident now referred to as the “sandwich incident”
• In the US, Treasury Secretary Bessent and Fed Chair Powell have convened Goldman Sachs and Bank of America executives before Congress to examine impacts on the financial system
• AWS, Apple, Palo Alto Networks, and CrowdStrike are participating in early-access program Project Glaswing; no Japanese companies are included
• Three actions for executives: ① run penetration tests using current AI, ② audit legacy assets, ③ elevate cybersecurity to a standing quarterly agenda item
• Dario Amodei puts general availability of Mythos at months to years out; Kajitani notes Anthropic is pulling ahead of OpenAI in the B2B market

2026-04-15 · Watch on YouTube

Japanese page: Claude Mythos：経営者が知るべきリスクと対策