SAST tools › AI SAST
AI SAST: what AI scanners find that pattern rules miss
AI SAST tools use a language model to find, triage or fix security bugs in source code. They catch auth and business-logic flaws that have no pattern for a rule to match, the kind a rule engine missed in the test below. They also cost more per scan and can vary between runs, so most teams add one beside a rule engine such as Semgrep or CodeQL rather than replacing it.
Four flaws a rule engine missed
Winding AI, which publishes this site and makes IonWarp, ran open-source Semgrep (1.185.0, the default ruleset) on the changed files of the 13 pull requests in IonWarp's security benchmark, which come from cal.com, Discourse, Grafana, Keycloak and Sentry. Semgrep matched 2 of the 18 known security issues: an unescaped request.referer in a script tag, and an unchecked open() call, the second under a different name (command injection where the key says server-side request forgery). It flagged nothing for any of these:
| Pull request | The flaw | Why no rule fires |
|---|---|---|
| cal.com 14740 | A guest blocklist is lowercased but the submitted emails aren't, so Victim@Example.com gets past it | Both lines are ordinary string calls; the bug is the mismatch between them |
| Grafana 103633 | A permission cache trusts cached grants but re-checks cached denials, so a revoked grant can keep working | Nothing is unsafe in isolation; the flaw is an asymmetry in the logic |
| Keycloak 37038 | canManage() accepts the VIEW scope as well as MANAGE, so view-only users get write access | The check exists; it checks for the wrong thing |
| Sentry 67876 | OAuth state uses a static pipeline signature instead of a per-request random value | A rule would need to know what this value should be |
On the same 13 pull requests, IonWarp's benchmark reports that Qodo found 11 of the 18 issues and IonWarp's default security reviewer 10, graded by a model against an answer key IonWarp chose. Both counts are small and unaudited: IonWarp's row is a September 2026 run of its model outside the hosted app, on an earlier prompt, and IonWarp's data file marks the security labels as disputed. The test shows the kind of flaw AI reads for, not a ranking of tools. Semgrep's other 15 findings, six of them SHA1 or MD5 hashes, weren't in the answer key, and nobody graded whether each one is worth fixing.
Three kinds of AI SAST
| Tool | Kind | What the AI does | 10 developers |
|---|---|---|---|
| Semgrep | Rules plus AI | AI detection, auto-triage and fixes on top of its rules; 20 AI credits per developer a month on Teams | $0 on Free (10 contributors, 10 repos); $300 on Teams |
| GitHub Code Security | Rules plus AI | CodeQL, plus AI Scan for pull requests in languages CodeQL doesn't cover, and Copilot Autofix | $300 + GitHub Team $40, plus AI credits at $0.01 |
| Checkmarx One Fusion | Rules plus AI | Adds frontier-model reasoning to Checkmarx SAST; Findings Analysis sorts true from false positives | Quote, on prepaid credits |
| Corgea | AI-native | AI SAST with logic and auth scanning, even on the Free plan | $390 (Growth) |
| ZeroPath | AI-native | AI SAST with business-logic and broken-auth detection, SCA and secrets | $1,000 + 10 × $60 = $1,600 |
| IonWarp | AI pull-request reviewer | Reads each pull request's changes for logic, auth and injection flaws; GitHub only | $149 (Max, 10 seats) |
Rules plus AI keeps the predictable scan and uses the model mostly to cut noise and suggest fixes; GitHub's AI Scan and Checkmarx Fusion also use it to find more. AI-native scanners start from the model and read whole repositories for flaws like the four above. AI reviewers read only what each pull request changes, which is cheap and fast, but they never look back at code already merged. Qodo and CodeRabbit work the same way; on IonWarp's benchmark Qodo found one more issue than IonWarp's default. Sonar also lists a new SonarQube Hunter Agent, "the AI security agent that hunts logic flaws", with no price on its plans page. The free way to try the AI-native approach is Datadog's open-source SAIST (Apache-2.0, in preview): it sends your code to an Anthropic, OpenAI or Google model on your own API key and writes a SARIF report.
What to check before you buy
- Repeatability. Checkmarx argues that AI-only scanners "can provide different results on each scan". Run any AI tool twice on the same commit during a trial and compare.
- Where your code goes. Semgrep sends part of the file with a finding to a model when you opt in, and says model vendors may not train on it. ZeroPath's Enterprise plan offers on-premises deployment and your own model keys; Corgea's Enterprise plan is single-tenant.
- Metered AI. GitHub's AI Scan and agentic autofix draw AI credits beyond the $30 license, and GitHub's docs advise a hard budget before turning them on widely. Semgrep and Checkmarx also meter AI in credits.
- The rest of SAST. An AI scanner that skips dependency scanning or secrets still leaves you needing those. Corgea and ZeroPath include both; IonWarp has no dependency scanning and checks secrets only in each pull request's diff.
Where IonWarp fits
IonWarp, from the company that publishes this site, is the AI-reviewer kind. It reads each GitHub pull request, sends the ones that touch auth, input, data access or secrets to its Security Review, and posts one comment. It is not a rules-based scanner: it doesn't scan the whole repository, has no dependency or CVE scanning, checks secrets only in each pull request's changes and works only on GitHub. Its own security benchmark has 13 pull requests. If your code is on GitHub and you already run Semgrep or CodeQL, it adds a model's read of risky pull requests for logic and auth flaws, from $49 a month for 5 seats, or $149 for 10. Visit IonWarp
Semgrep pricing · Corgea pricing · ZeroPath pricing
Questions
What is AI SAST?
Static application security testing that uses a language model to find vulnerabilities, sort real findings from false ones, or write fixes. Some tools add a model to a rule engine; others, such as Corgea and ZeroPath, are built around one.
Is AI SAST better than traditional SAST?
It finds different things. Rules catch known patterns the same way on every run; models catch auth and business-logic flaws that have no pattern, but can vary between runs and cost more per scan. Most teams keep a rule engine and add AI on top.
Does AI SAST send my code to a model provider?
Usually, yes, for the AI part. Semgrep says that when you opt in, part of the file with a finding goes to a model, and vendors may not train on it. ZeroPath's Enterprise plan lets you bring your own model keys or run it on-premises. Ask each vendor which provider sees your code.
Related
- Open-source SAST tools: the free rule engines to keep as your baseline, and Datadog SAIST.
- Checkmarx alternatives: where Fusion's quote fits against published prices.
About this page
Published by Winding AI, Inc., the company that makes IonWarp, an AI pull-request reviewer for GitHub. We earn nothing from links to other tools; links to IonWarp carry a tag so IonWarp can count visits from this site. Visits are counted with Google Analytics. If a price or fact here is wrong, write to support@winding.ai.