OpenAI isn't far behind Mythos' hacking powers
The head start that cyber defenders were promised when Mythos was unveiled last month is disappearing faster than expected.
TL;DR
- OpenAI's GPT-5.5 model shows nearly equivalent performance to Anthropic's Mythos Preview in finding and exploiting software bugs.
- GPT-5.5 completed a 32-step simulated corporate cyberattack in 2 out of 10 test runs, compared to Mythos' 3 out of 10.
- GPT-5.5 also outperformed Mythos on capture-the-flag tasks testing vulnerability identification, incident reverse-engineering, and web application exploitation.
- The rapid advancement challenges Anthropic's initial estimate of a 6-18 month lead time for similar AI cyber capabilities.
- Access to both Mythos and advanced GPT models is limited to select organizations and vetted cyber defenders.
- The White House has reportedly urged Anthropic to limit Mythos access due to national security concerns.
- OpenAI is providing access to cyber-permissible versions of GPT-5.4 and 5.5 to governments and allies.