AI security testing: SAST, DAST, IAST in the LLM era.

The three categories, briefly

SAST (Static Application Security Testing) analyses source code without executing it. Strengths: comprehensive coverage of the code, no runtime needed, catches issues early. Weaknesses: high false-positive rate, blind to runtime context, requires source.

DAST (Dynamic Application Security Testing) probes a running application from the outside, simulating an attacker. Strengths: catches issues that only appear at runtime, no source needed. Weaknesses: coverage depends on traffic generation, slower to run, blind to code-level issues.

IAST (Interactive Application Security Testing) instruments the running application and observes real execution paths during testing. Strengths: low false-positive rate, deep context, catches both code and runtime issues. Weaknesses: requires instrumentation, can affect performance, limited language support.

Each catches a different class of issue. A mature security testing programme uses all three; an early-stage programme typically starts with SAST in CI and adds DAST or IAST as the application matures.

Where AI augments SAST

The dominant SAST tools (Snyk, GitHub CodeQL, SonarQube, Checkmarx, Veracode, Semgrep) have all added AI features in the past two years. The common patterns:

False-positive triage. Traditional SAST tools flag thousands of issues in a non-trivial codebase, the majority of which are noise. AI-classified triage prioritises real issues over noise, reducing the security team's review load. Snyk DeepCode AI (snyk.io/platform/deepcode-ai) and GitHub's CodeQL with AI-augmented review are published examples.

Fix suggestion. GitHub Copilot Autofix proposes patches for vulnerabilities flagged by CodeQL. Developers review the suggested patch in the PR and accept or reject; the published value is faster mean-time-to-fix on known vulnerability classes.

Custom rule generation. Semgrep and similar tools let engineers describe a vulnerability pattern in natural language and generate a custom rule from the description. This shortens the cycle from "we found a new bug class" to "we catch it across the codebase."

Where AI augments DAST

DAST is harder to augment with AI because the input is the running application's response surface rather than source code. The published patterns:

Smart crawling. Traditional DAST tools crawl the application by following links and submitting forms. AI-augmented crawlers reason about the application's state, navigate complex multi-step flows, and reach pages that simple crawlers miss. The coverage improvement is real and material.

Payload generation. Generating sensible fuzz payloads (SQL injection, XSS, command injection variants) is a task LLMs do well given a description of the application's technology stack. This expands DAST coverage without requiring custom payload curation.

Result triage. DAST tools also flag many false positives; AI-classified triage helps here as it does in SAST.

Where AI augments IAST

IAST's lower false-positive rate means less triage burden, so AI value is concentrated in two places: extending language coverage (interpreting instrumentation output in languages that vendor IAST agents do not natively support) and correlating findings with SAST and DAST results to produce a unified vulnerability view. Contrast Security and similar IAST vendors publish these capabilities.

What AI does not help with

Threat modelling. Deciding what the threats are, who the threat actors are, what assets matter, and what countermeasures justify investment is design work informed by business context. AI can produce a draft threat model from a system description but the senior judgement work remains human.

Incident response. When a vulnerability is found in production, the response (containment, remediation, communication, lessons learned) is a coordination and judgement exercise. AI helps with information retrieval but does not run the incident.

Pen testing at the senior level. Adversarial, context-aware, goal-directed pen testing is harder for AI than the marketing material suggests. AI-augmented security scanners catch the obvious; the hard findings remain a senior pen tester's territory.

Compliance contexts

For SOC 2-audited codebases, the security testing tooling must satisfy the auditor that vulnerabilities are detected and remediated. AI-augmented tools are acceptable as long as the audit trail (what was scanned, what was found, what was fixed) is complete and the remediation is timely. Most major SAST and DAST tools have SOC 2-friendly audit logging built in.

For HIPAA-handling code, the additional concern is whether the AI tool sends source code or runtime traces to a third-party model. Self-hosted SAST options (CodeQL on a self-hosted GitHub Enterprise, Semgrep CLI) avoid the third-party-model question. Cloud-based AI security tools typically have data-handling policies that exclude customer source from model training but buyers should confirm rather than assume.

For PCI-DSS scope, the same considerations apply with the additional requirement that the tooling supports the specific PCI-required scans (typically quarterly DAST against in-scope assets). Most enterprise DAST tools are PCI-DSS aware.

Where this category is going

The realistic 2026 to 2027 trajectory: AI-augmented SAST will continue eating false-positive noise; AI-augmented DAST will continue extending crawl coverage; IAST will continue being the highest-precision category for teams that can deploy the instrumentation. The senior security engineering role becomes more leveraged rather than less needed; the engineer who can interpret AI findings, prioritise the queue, and run incident response is more valuable than the engineer who hand-triaged a SAST queue.

Frequently asked questions

Does AI replace a security engineer?

No. AI augments security tooling by reducing false-positive volume and suggesting fixes, but the security engineer's role of threat modelling, architectural review, and incident triage is not replaced. The most realistic 2026 framing is that AI removes some grunt work and surfaces some non-obvious patterns, while the senior judgement role remains human.

What is the difference between SAST, DAST, and IAST?

SAST (static) analyses source code without executing it. DAST (dynamic) probes the running application from the outside. IAST (interactive) instruments the running application and observes real execution paths. Each catches a different class of vulnerability and each has different false-positive profiles. The categories complement rather than replace each other.

Is Copilot Autofix trustworthy?

GitHub Copilot Autofix proposes fixes for vulnerabilities flagged by CodeQL. The proposed fixes are useful starting points; teams should review and test them before merge rather than auto-merge. The accuracy is high for well-known vulnerability patterns and lower for context-specific issues. Treating it as a faster developer is closer to the right calibration than treating it as an oracle.

How does AI handle zero-day-style novel vulnerabilities?

AI-augmented SAST and DAST work best on known vulnerability classes (OWASP Top 10, CWE-listed patterns). Novel vulnerabilities specific to an application's logic are harder; AI may surface candidates but a security review is still needed. Bug bounty programmes and threat-modelling work remain meaningful for novel-vulnerability discovery.

Does AI testing replace penetration testing?

No. Pen testing is adversarial, context-aware, and goal-directed work that AI does not currently replicate at the senior level. Automated AI-augmented security tools complement pen testing by catching obvious vulnerabilities before the pen tester arrives, which lets the pen tester focus on the harder findings.