Microsoft confie 90 % de sa chasse aux failles à un mini-modèle

Microsoft has unveiled a new compact model, MAI-Cyber-1-Flash, which is derived from its in-house reasoning model MAI-Thinking-1, designed to identify software vulnerabilities. When combined with GPT-5.4 in the multi-agent harness MDASH, it achieves a remarkable 95.95% on the CyberGym benchmark, up from 88.45% in May. The model handles up to 90% of common security tasks and halves the costs of production configuration.

The platform, named Project Perception, featuring red, blue, and green agents, is set to enter public preview on August 3 in Microsoft Defender, a pay-per-use service called Security Compute Units. This dual announcement features a tiny model trained for a single task and a platform of agents aimed at maintaining a continuous defensive posture.

The decision to move away from sending every line of code to a pricey generalist and towards a cheaper specialist in parallel with a generalist model like MAI-Thinking-1 is a strategic shift in the industry. Microsoft's AI models specialize in repetitive tasks, while generalist models like GPT-5.4 handle more complex cases. The company argues that this approach results in a 50% reduction in costs compared to the current MDASH configuration in production.

Competitors like Google DeepMind and Anthropic have taken different approaches, with DeepMind using CodeMender on Gemini models and Anthropic building Claude Security on generalist models. The vertical specialization of models is a key trend in the cybersecurity market, as it allows for efficient training and evaluation due to the availability of proprietary data, a large volume of nearly identical tasks, and a mechanical judge.

In contrast, other fields often rely on human judgment, leading to slower progress. The success of specialized models in cybersecurity is attributed to their ability to self-correct through proof, making the cybersecurity sector a prime market for smaller, vertical models, beyond just the hype. Microsoft's MDASH pipeline orchestrates over a hundred specialized agents in several stages: preparation, scanning, contradictory debate among agents, de-duplication, and exploit proofing.

On the CyberGym benchmark for real-world vulnerability detection in open-source code, the updated model improved from 88.45% in May to 95.95%. This score outperforms Anthropic's Claude Mythos by a dozen points. The pipeline's most critical steps, contradictory debate and exploit proofing, are designed to eliminate false positives.

These stages address the main adversary, noise, which is not quantified in public metrics. The question arises when an agent acts without review. A false alert in a dashboard costs minutes of an analyst's time, while an automatically applied fix affects production code. Setting the tolerance threshold for alerts means deciding whether a loop can close without human intervention.

Microsoft's Project Perception distributes three roles: red agents simulate attacks like an attacker looking for an open door, blue agents detect threats and prioritize them by severity, and green agents apply fixes and close breaches. Impactful actions remain subject to human validation, while Microsoft defends this line of separation, stating that the "high impact" workload is defined by the platform, not the user.

The economic formula introduces a second tension: consumption-based billing based on Security Compute Units, which depends on the intensity of tasks executed. A zealous agent that increases scans becomes a budget line item, and cost optimization ultimately dictates the depth of analysis. Users are encouraged to test the model on their own codebases to measure the number of alerts generated, rather than relying solely on benchmark tests.