OpenAI Cyber Model Outscores Mythos, WH Silent
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI debuted GPT-5.5-Cyber on Monday, scoring 85.6% on the CyberGym benchmark for AI agents reproducing known software vulnerabilities — beating Anthropic's Mythos 5 (83.8%) on the same evaluation.
- Anthropic remains bound by Trump administration export controls that bar foreign nationals from using its models, with reports citing personality clashes between the company and the White House as the trigger for the directive.
- The White House and OpenAI declined to comment to Axios on why the administration moved against Anthropic's comparably capable cybersecurity models while letting OpenAI's release pass without pushback.
- Mythos access was a major talking point at last week's G7 Summit, where AI CEOs negotiated global AI rules with the Trump administration — underscoring the geopolitical stakes of the export restrictions.
- OpenAI expanded cybersecurity partnerships with organizations in Australia, Canada, France, Germany, Japan, Poland, South Korea, and the EU alongside the model release.
Why it matters: OpenAI can now deepen cybersecurity partnerships across eight allied markets while Anthropic remains barred from foreign-national use of its models — and the White House's refusal to explain the differential treatment, combined with reports that 'personality clashes' drove the export directive, suggests the action was political rather than capability-based, since a 1.8-point CyberGym gap is a thin pretext for singling out one firm.
