Microsoft AI Releases MAI-Cyber-1-Flash: A 5B-Active-Parameter Cyber Model
A new cybersecurity specialized model hitting 95.95% on CyberGym while cutting operating costs in half.
Stock photo for illustration only, not from the actual event
- MAI-Cyber-1-Flash features 137B total parameters with 5B active
- Pushes MDASH score on CyberGym to a remarkable 95.95%
- Delivers a 50% cost reduction by handling 90% of routine tasks
- Developed by Microsoft ACS and DARPA AI Cyber Challenge winners
Microsoft AI has officially introduced MAI-Cyber-1-Flash, a specialized artificial intelligence model tailored for cybersecurity tasks. Derived from the MAI-Thinking-1 lineage, it serves as a cybersecurity-focused fine-tune of MAI-Code-1-Flash, the lightweight agentic coding model currently integrated into GitHub Copilot and VS Code. The model accepts text-only inputs and outputs.
Under the hood, MAI-Cyber-1-Flash utilizes a transformer architecture equipped with self-attention and sparse Mixture-of-Experts layers. It packs a total of 137 billion parameters, out of which 5 billion are active, alongside an extensive context length of 256k tokens to thoroughly analyze complex codebases.
Stock photo for illustration only, not from the actual event
Microsoft evaluated the model using its MDASH harness—combining MAI-Cyber-1-Flash and GPT-5.4—on CyberGym, a public benchmark suite comprising 1,507 real-world vulnerability reproduction tasks drawn from 188 OSS-Fuzz projects. Running at the default level 1 configuration, the setup achieved a score of 95.95%, outperforming Anthropic's Mythos by roughly 12 points.
MDASH orchestrates over 100 specialized agents across five distinct stages: Prepare, Scan, Validate, Dedupe, and Prove. Auditor agents flag potential issues while debater agents argue exploitability using disagreements as signals, ultimately triggering inputs with ASan for C and C++ targets.
Deploying a smaller, highly optimized 5B-active model to shoulder 90% of the workload highlights an efficient routing strategy in frontier AI development. By reserving massive frontier models like GPT-5.4 exclusively for the most demanding 10% of tasks, engineering teams can dramatically curb operational overhead without sacrificing accuracy.
To keep frontier model expenses under control at scale, MAI-Cyber-1-Flash processes up to 90% of MDASH tasks and escalates the hardest 10% to GPT-5.4. This routing mechanism results in a 50% cost reduction compared to previous configurations involving GPT-5.4, 5.4 mini, and 5.3 codex.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment