Cybersecurity AI is splitting into four bets, and buyers still lack a common scoreboard

Cybersecurity companies are taking sharply different routes into AI, from purpose-built defender models to frontier systems wired into existing security platforms. The practical problem is that buyers are being handed metrics measured in different units, making the next big security purchase rather harder to compare than the marketing suggests.

Watch Desk analysis

What happened

VentureBeat compares four approaches from CrowdStrike, Google, Palo Alto Networks and Microsoft. CrowdStrike is building specialised models trained on security data, Google has released a security-tuned Gemini variant, Palo Alto is integrating third-party frontier models into Cortex, and Microsoft is combining security-specific tooling with Azure OpenAI services.

The article says Amazon CISO CJ Moses described a honeypot capturing an AI agent that completed a full attack in 12 minutes and 42 seconds. CrowdStrike, meanwhile, reported an 89% year-on-year increase in attacks it classified as AI-enabled. Those are attributed accounts, not a shared independent test, but they explain why the industry is treating autonomous security work as an immediate architecture question rather than a distant demo.

The vendors are also measuring different things. CrowdStrike reports higher detection and faster remediation for SafeMind. Google reports scores for vulnerability discovery and patching. Microsoft reports analyst productivity in a controlled trial, including a claim that phishing triage users identified up to 6.5 times as many true positives per analyst minute. VentureBeat notes that much of Microsoft's reported gain came from queue management and automatic resolution of apparently benign reports, rather than verdict quality alone.

Read VentureBeat's comparison of the four security AI bets.

Why it matters

A detection uplift, a benchmark patching score and analyst time saved are not interchangeable currencies. Treating them as one league table would be like comparing a fire engine by its top speed, its hose pressure and the number of sandwiches eaten by the crew. All may matter, but they answer different questions.

For buyers, the useful test is architectural as much as numerical: who owns the security-specific training, where customer data goes, how model changes are controlled, and how painful it would be to switch platforms later. A system that looks impressive in a vendor's preferred test may still be a poor fit for an organisation with different telemetry, permissions or tolerance for automated action.

Our read

The strongest takeaway is not that one vendor has won. It is that the market needs a common evaluation frame before security teams commit to multi-year AI systems. Buyers should ask each supplier to define the task, disclose the baseline, show what was automated and explain what happens when the model is wrong.

The article's vendor performance figures remain company-reported, and the supplied evidence does not provide a common independent benchmark. That is not a footnote. It is the central purchasing problem.

What to watch

• Whether security vendors publish comparable task-level results rather than isolated headline percentages.
• How each system handles customer data, model updates and third-party inference.
• Whether autonomous agents can be limited by identity, permissions and tool access before they act.
• Independent testing of claimed detection, patching and remediation improvements.

Discussion spark: Should security teams prioritise a specialised defender model with narrower control, or a general frontier model with broader capability and a stronger platform ecosystem?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.