Science Weekly Mythos: are fears over new AI model panic or PR?
27 snips
Apr 21, 2026 
Aisha Down, a Guardian reporter who investigates AI companies and policy, breaks down Anthropic’s Mythos claims. She discusses alleged autonomous exploit-finding, who gets access through Project Glasswing, and expert skepticism. Aisha also explores whether the messaging was political PR or a genuine safety warning and what that means for sensible regulation.
AI Snips
Chapters
Transcript
Episode notes
Mythos Claims Versus Missing Technical Evidence
- Mythos reportedly found many zero-day vulnerabilities and can chain exploits to perform autonomous attacks.
- Anthropic announced the results in a blog post without technical papers, so experts lack details on cost, prompting effort, and false positive rates.
Project Glasswing Limited Access List
- Anthropic limited Mythos access to Project Glasswing partners including AWS, Apple, Broadcom, Cisco, CrowdStrike and JPMorgan Chase.
- Aisha notes select big tech, banks, and security researchers got access while no papers were released.
Similar Models Can Replicate Mythos Findings
- Other open-source models reproduced Mythos-style exploit findings in tests, suggesting the capability sits near the frontier rather than being unique.
- Former Anthropic researchers at AI Sile tested other models and replicated results eight out of eight times.