Microsoft Launches MAI-Cyber-1-Flash, Beats Anthropic's Mythos on Vulnerability-Hunting Benchmark
Microsoft's first dedicated cybersecurity model scores 95.95% on the CyberGym vulnerability-reproduction benchmark, ahead of Anthropic's Mythos, and ships alongside an agentic security platform called Project Perception.
Microsoft announced MAI-Cyber-1-Flash on July 27, its first purpose-built model for finding and fixing software vulnerabilities, alongside an agentic platform called Project Perception that deploys red, blue, and green agent teams to simulate attacks, triage bugs, and ship fixes automatically.
- Scores 95.95% on CyberGym, a benchmark of 1,507 vulnerability-reproduction tasks, up from 88.45% for the prior MDASH harness version
- Anthropic's Mythos scored roughly 84% on the same benchmark
- Cuts the cost of running the harness by about half versus the previous version
- Handles about 90% of the workload itself, routing the hardest 10% of tasks to OpenAI's GPT-5.4
- Priced on a consumption basis in 'Security Compute Units'; public preview begins August 3 inside Microsoft Defender and via Azure AI Foundry
The launch puts Microsoft in direct competition with Google's Gemini 3.5 Flash Cyber and Anthropic's Mythos, both released earlier this year, as major labs race to ship specialized security models rather than relying on general-purpose flagships for vulnerability work.