AI vulnerability detection is becoming a routing problem, not a tooling upgrade
AI vulnerability detection is the use of specialized cybersecurity AI models and agents to scan code, localize weaknesses, confirm exploitability, and propose fixes, shifting security work from manual bug hunting toward automated bug finding with human review and policy-driven control over which systems act on which risks. The most important shift now is not that vendors can scan more code; it is that they are routing different security tasks to different models so they can cut cost without giving up accuracy. Microsoft’s MDASH does this openly: its MAI-Cyber-1-Flash model is tuned to handle up to 90% of routine vulnerability-identification tasks, while GPT-5.4 is reserved for the hardest 10%. That routing strategy, Microsoft says, delivers a 95.95% CyberGym score at 50% of the cost of its previous MDASH model mix. In other words, AI bug hunters are not one big brain; they are traffic controllers deciding which brain to pay for on each task.
Microsoft: MAI-Cyber-1-Flash turns accuracy into a cost game
Microsoft’s play is unapologetically systemic: MAI-Cyber-1-Flash is not sold as a standalone chatbot but as the core of MDASH, its multi-model vulnerability identification and remediation system. The model itself is a sparse mixture-of-experts transformer with 137 billion parameters, five billion active per call, and a 256,000‑token context window. That scale matters less than where it sits in the pipeline. By designing MAI-Cyber-1-Flash to take up to 90% of MDASH tasks and routing only the hardest 10% to GPT‑5.4, Microsoft reports that it replaced 80% of its previous internal models and raised its CyberGym score from 88.4% to 95.95% at half the cost of its earlier GPT‑5.4, GPT‑5.4 mini, and GPT‑5.3 Codex combination. "Taken together, the disclosed design points to routing as the central technical claim" in MDASH’s performance gains. The signal for security leaders is clear: the competitive frontier is AI vulnerability detection architectures that spend frontier-model compute only where it truly helps.
Google: CodeMender shifts from code security scanning to AI-led remediation
Where Microsoft focuses on system-level routing, Google’s CodeMender aims to change what a single AI agent can own end-to-end. In preview, CodeMender runs as a security AI agent that scans repositories for vulnerabilities that static tools often miss, including memory corruption, injection flaws, web security issues, cryptographic weaknesses, and insecure data handling. It then tries to confirm which findings matter by building proof-of-concept exploits and running them in a customer-controlled sandbox, cutting down on false positives by testing exploitability rather than trusting patterns. Once a flaw is confirmed, the agent writes a patch and has a model judge review it for regressions before handing the result to a human reviewer. The patch reaches developers as a code diff, and no change reaches a repository without manual approval. Google argues that this can move organizations from passive code security scanning to automated code remediation and reduce zero-day risk, while still keeping humans in charge of merges.
Cisco: Antares proves local AI triage is powerful but not a verdict
Cisco’s Antares family takes a different bet: that security teams want AI vulnerability detection to live inside their own environment, even if that means accepting limited autonomy. Released on July 21, 2026, the open‑weight Antares‑350M and Antares‑1B models are designed to shorten one of application security’s slowest jobs—finding where a known weakness may exist inside a large codebase. Antares works as a terminal-based code-search agent rather than a stand-alone vulnerability scanner, using commands like grep, find, and cat against a read-only repository to surface files that may contain instances of a specified CWE. Antares‑350M offers a 32,000‑token context window for constrained deployments, while Antares‑1B extends that to 128,000 tokens on a single GPU. Local operation is the headline: Antares can localize potentially vulnerable files without relying on a cloud AI service, helping organizations examine proprietary code on-premises. But Cisco is explicit: Antares does not confirm vulnerabilities, assign severity, or generate fixes, and its findings still need human verification before they influence remediation decisions.
Security teams now manage AI workflows, not checklists
Taken together, Microsoft, Google, and Cisco show the unavoidable trade-off: automated bug finding will not remove humans from the loop, but it will change where they work. MAI‑Cyber‑1‑Flash’s model card warns that generated text and code may be inaccurate or incomplete and should be reviewed before consequential use, even as MDASH automates more of the path from description to proof-of-concept exploit. CodeMender pushes further into automation, yet still enforces that every patch is human-approved before entering a repository. Antares is even more conservative: it should provide advisory leads rather than block builds, approve fixes, or replace static analysis and software composition analysis. The practical result is that security teams must adapt to higher volumes of machine-generated findings, triaging AI output instead of hand-running scans. Model routing and task specialization are becoming key cost-reduction strategies, but the work now is to increase review capacity, redesign workflows, and decide which AI verdicts are trusted enough to act on and which remain hints.
