Static application security testing (SAST) has been a staple of AppSec for two decades. It finds real bugs, and it has well-known limits: it matches patterns and tracks data flows, but it does not understand what code is supposed to do. That leaves whole categories, such as logic flaws, authorisation mistakes and subtle memory-safety issues, largely to human reviewers, penetration testers and fuzzers.
This year, AI has started to change that picture.
What happened in 2024
- Google's Big Sleep (a collaboration between Project Zero and DeepMind) announced on 1 November that its AI agent had found an exploitable stack buffer underflow in SQLite, a heavily tested and fuzzed codebase. The bug was fixed before it reached an official release.
- Later in November, Google reported that AI-generated fuzz targets in OSS-Fuzz had helped find 26 new vulnerabilities in open-source projects, including one in OpenSSL (CVE-2024-9143) that had likely been present for around two decades.
- Other teams released LLM-based tools that trace user input through Python projects to find exploitable vulnerabilities, reporting previously unknown bugs in popular open-source projects.
- DARPA's AI Cyber Challenge semifinals in August showed automated systems finding and patching vulnerabilities in real open-source code.
The common thread: AI systems finding real, previously unknown vulnerabilities, some in code that traditional tools had examined for years.
How AI-driven testing differs from SAST
| Traditional SAST | AI-assisted security testing | |
|---|---|---|
| Approach | Rules, patterns, data-flow analysis | Reads and reasons about code like a reviewer |
| Strength | Consistent, fast, exhaustive for known patterns | Context, intent, novel and logic bugs |
| Weakness | False positives, blind to logic flaws | Non-deterministic, costly, can hallucinate |
| Output | Findings with rule IDs | Hypotheses, often with proof-of-concept inputs |
| Verification | Human triage | Execution: tests, sanitisers, crash reproduction |
The most promising systems combine the two: they use models to hypothesise vulnerabilities and generate inputs, then use execution (running code, sanitisers and debuggers) to confirm them, which filters out hallucinated findings.
flowchart LR C[Codebase] --> H[AI agent reads code,<br/>forms hypotheses] H --> G[Generates test inputs<br/>or fuzz harnesses] G --> X[Execute with sanitisers<br/>and debuggers] X -->|crash / violation| V[Confirmed vulnerability<br/>+ reproducer] X -->|no effect| H V --> T[Human triage & fix]
What this means for AppSec programmes
AI augments, it does not replace
SAST, SCA, DAST and fuzzing remain essential. They are deterministic, cheap to run on every commit and exhaustive for what they cover. AI testing adds coverage for what they miss, and is likely to run less frequently on the most critical code.
Expect more findings
As AI-driven discovery improves, the number of genuine vulnerabilities found in both your code and your dependencies will rise. The constraint will shift from detection to triage and remediation capacity. Invest now in:
- fast, reliable triage processes,
- clear ownership and remediation SLAs, and
- the ability to ship patches quickly.
Point AI at your most critical code first
Parsers, authentication, payment logic, cryptographic code and anything processing untrusted input are where AI-assisted review and fuzzing earn their cost.
Insist on reproducibility
For AI-generated findings, require a reproducer (an input, a test or a proof of concept) before investing engineering time. Execution-verified findings are far more trustworthy than model assertions.
Assume attackers have it too
The same techniques are available to attackers. The window between a vulnerability being findable and being found is shrinking. Patch cadence, dependency hygiene and defence in depth matter more.
A planning checklist for 2025
- Keep deterministic scanners in every pipeline; tune for signal
- Pilot AI-assisted review or fuzzing on the most critical components
- Require reproducers for AI findings
- Measure triage and remediation capacity; plan for growth
- Add AI-found vulnerability classes to secure coding training
- Review your patch pipeline: how fast can you ship a fix to production?
The bottom line
Application security testing is entering a new phase. Pattern-based tools will remain the foundation, and AI systems that read, reason and verify by execution are beginning to find what those tools miss. The organisations that benefit will be the ones ready to handle more findings, faster, with the triage and remediation discipline to match.
Our secure code review combines AI-assisted analysis with expert validation, and our application security testing team helps organisations adopt new testing approaches without drowning in findings.