Microsoft launches AI cybersecurity model scoring 95.95% on benchmark

Microsoft’s new MAI‑Cyber‑1‑Flash model, paired with the MDASH platform and OpenAI’s GPT‑5.4, achieved a 95.95 % score on the CyberGym benchmark, a result Microsoft says cuts security‑compute costs by about half compared with its existing configuration [2].

Model architecture and rollout

MAI‑Cyber‑1‑Flash is a compact, in‑house security model designed to handle roughly 90 % of vulnerability queries, delegating the remaining 10 % to the larger GPT‑5.4 model for especially tough problems [2]. The combined system—dubbed Project Perception—enters public preview on August 3 and will be integrated gradually across Microsoft Defender and other security products [1]. Microsoft frames the 90/10 split as a way to deliver “near‑frontier” performance while keeping token consumption low, which it argues is the primary cost driver for enterprise AI workloads.

Competitive context

The 95.95 % CyberGym result outperforms rival configurations from Google, OpenAI and Anthropic by more than 10 percentage points, according to Microsoft’s own benchmarks [2]. By contrast, Anthropic’s Claude Mythos 5 scored 84 % on the same test, a 12‑point gap to Microsoft’s combined system [1]. The claim of “almost 50 %” cost savings versus Microsoft’s own MDASH setup hinges on the consumption‑based pricing model, where each security compute unit (SCU) incurs a fee proportional to workload volume [1].

Cost‑focused positioning

Microsoft’s CEO of AI, Mustafa Suleyman, emphasized that token costs, not raw model quality, are the barrier to enterprise adoption. He noted that the new architecture reduces reliance on the expensive GPT‑5.4 tier, leveraging the smaller MAI‑Cyber‑1‑Flash for the bulk of work and reserving the larger model for edge cases [2]. This mirrors a broader market trend toward “cheaper‑enough” models, as seen in recent releases from xAI and Chinese AI firms that aim to undercut frontier‑model pricing.

What to watch

• August 3 – Public preview launch of Project Perception across Microsoft Defender.
• Enterprise rollout – Pace of integration into the broader Microsoft Security suite.
• Competitor response – Potential updates from Anthropic, Google or OpenAI to address the benchmark gap.

Microsoft’s claim of a sub‑50 % cost advantage and a near‑frontier benchmark score could reshape enterprise AI security purchasing, but the true impact will depend on how quickly customers adopt the new system and whether rivals can match its performance‑cost balance.