How does AI improve vulnerability detection and where does it fall short?

AI improves vulnerability detection mainly in three places: reading code for flaw patterns that rules miss, triaging scanner findings, and judging whether a flaw is reachable. It can also invent findings or miss real ones, so its output needs verification by tests, rules or a person before it drives action.

Where does AI help?

Pattern reading. Static analysis rules catch patterns someone wrote a rule for. A model can read a function and notice that a value from a request reaches a query without a check, even in an unfamiliar framework. It extends coverage, especially for custom code.

Triage. Scanners produce many findings, and most teams cannot fix them all. A model can read the finding and the surrounding code, and say that the flagged function is test only code or that the input is already validated upstream. That removes noise.

Exploitability. Whether a vulnerable library function is called by your code is a question about program flow. Models and analysis tools can trace this and rank findings accordingly.

Explanation and fixes. A model can write a plain description of the flaw and propose a patch, which speeds the developer's work.

Where does it fall short?

Models produce confident wrong statements. In detection that shows up as false positives, which waste time, and false negatives, which are worse because nobody looks again. Results can vary between runs on the same code. Long codebases exceed what a model can hold in view, so findings may depend on which files were included. And a model proposing a fix can introduce a new flaw.

It also inherits the attack surface of AI systems. NIST AI 100-2 describes adversarial attacks on machine learning, including manipulation of inputs at inference time. A comment in source code that tells an AI reviewer to ignore a flaw is a simple example of text addressed to the tool, so treat code as untrusted input to any AI reviewer.

How should you combine AI with existing methods?

Use AI as an additional signal, not as the sole judge. Deterministic tools give repeatable results and clear rules. AI adds breadth and context. The practical approach is to run both, require a second form of evidence for high severity findings, such as a failing test or a traced path, and let a person approve fixes.

For prioritisation, pair findings with outside data. CISA's Known Exploited Vulnerabilities Catalog lists vulnerabilities with evidence of active exploitation. The FIRST Exploit Prediction Scoring System estimates the likelihood that a vulnerability will be exploited. Neither depends on a model reading your code, which makes them a useful cross check.

NIST SP 800-218, the Secure Software Development Framework, asks organisations to review and analyse code and to test executable code for vulnerabilities. It does not prescribe a technique, so AI assisted methods can serve those practices if the results are verified.

A constructed example

A team's static scanner reports forty findings in a service. An AI triage step reads each in context. It marks twenty three as unreachable or already guarded, with a one line reason each. A developer spot checks five of those and agrees with four. The fifth is a real flaw the model dismissed because of a validation function that only covered one of two code paths.

The team keeps the triage but changes the rule: dismissals of high severity findings need a human sample, and a test must exist for any flaw marked guarded.

How to check or apply this

1. Measure first: run your existing scanner and record findings, false positives and time spent.
2. Add AI triage or detection in a pilot on one service, and compare against the baseline.
3. Check a random sample of dismissed and accepted findings by hand each week.
4. Require a test or traced path for high severity results.
5. Treat source text and comments as untrusted input to the AI.
6. Record model and version, since results change with them.
7. Prioritise by exploitation evidence from KEV and likelihood from EPSS, not by model confidence.
8. Keep fixes behind review and CI.

Safeguard runs reachability analysis to rank findings by whether your code reaches the vulnerable function, supports zero day discovery, and can open fix pull requests, which stay off until you enable them.

Quick check

Sources

• NIST SP 800-218, Secure Software Development Framework
• NIST AI 100-2 E2025, Adversarial Machine Learning
• CISA Known Exploited Vulnerabilities Catalog
• FIRST, Exploit Prediction Scoring System

Last reviewed: 18 October 2026.