GPT-5.5 matches heavily hyped Mythos Preview in new cybersecurity tests
Summary
New tests from the UK’s AI Security Institute show that OpenAI’s GPT-5.5 performs at a similar level to Anthropic’s Mythos Preview in cybersecurity tasks. Both AI models demonstrated strong abilities in various hacking and code challenges, but they still struggle with the most complex simulations like power plant disruptions.Key Facts
- GPT-5.5 and Mythos Preview were tested by the UK’s AI Security Institute on 95 cybersecurity challenges.
- GPT-5.5 passed about 71.4% of the hardest “Expert” tasks, slightly more than Mythos Preview’s 68.6%.
- GPT-5.5 completed a difficult task of decoding Rust code in just over 10 minutes without help, costing $1.73 in API use.
- Both models showed progress on a 32-step data attack simulation, which no earlier AI had passed.
- Neither GPT-5.5 nor Mythos Preview could solve a complex test simulating a power plant control attack.
- OpenAI’s CEO Sam Altman criticized marketing that promotes AI threats as extremely dangerous to drive sales.
- OpenAI has a Trusted Access program for security experts to study advanced AI models responsibly.
- The limited release of GPT-5.5-Cyber is planned to go only to trusted cyber defenders soon.
Read the Full Article
This is a fact-based summary from The Actual News. Click below to read the complete story directly from the original source.