What is AI SAST?

What it is

Aikido's explainer on "AI SAST" — static analysis where an LLM reasons about code rather than matching patterns. The summary captures the AI-native vs. AI-augmented split and the three-tool argument; what it omits is the buyer-facing diagnostics, the cost figures, and the concrete exploit-chain examples that make the argument operational.

How to tell what a vendor is actually selling

The page gives a listening test, useful when evaluating a pitch:

• Pitch is about reducing false positives, triaging, explaining results, or suggesting fixes → detection is still rule-based; the AI is an assistant. This is AI-augmented SAST (same role as traditional SAST).
• Pitch is about reasoning across the codebase, catching IDORs, finding business-logic flaws, or following data flow across services → an AI is doing the detection. This is AI-native SAST.

The page's stated position: AI augmentation on deterministic scanners "should be treated as a requirement instead of a nice-to-have," so it is not a differentiator — only reasoning-based detection creates a new category.

Cost and cadence — the numbers

The concrete figures the summary leaves out:

• Full-repo AI-native run: roughly half an hour and a couple hundred dollars per pass. Scheduled for high-stakes changes and release audits, not every commit.
• Diff-scoped run (Deep PR Review): scoping the same reasoning to a PR's diff drops cost enough to run on every PR.
• Price ordering: AI SAST sits between rules-based SAST and AI pentesting.
• Rule of thumb (whole codebase): SAST always, AI SAST often, AI pentest here and there — plus PR-scoped AI SAST on every PR.

What it catches that pattern-matching can't

The mechanism example worth carrying: a chain of individually low-severity findings — an information disclosure, a weakly-scoped session token, and a permissive CORS policy — that an AI connects into a full account takeover, because it reasons across findings rather than firing per-rule. Named classes: IDORs/broken access control (trace request from route handler through DB query, notice nothing enforces ownership), privilege escalation across service boundaries, business-logic bypasses (e.g., a subscription-tier check missing on one endpoint).

One structural advantage the summary skips: language coverage. Rules must be written per-language; an LLM is "literate" across languages, so AI SAST covers codebases in uncommon or mixed languages without per-language rule development.

Caveats (the page's own)

AI SAST reads code without running it, so it produces more false positives than a live pentest — it can flag something that looks vulnerable but never fires at runtime, and it can't validate exploitability. That is the explicit trade for its wider net and its ability to reach code paths a live test can't safely exercise. This is a vendor post; the closing maps everything onto Aikido's Code Security Audit and Deep PR Review.