Post

Why Don’t We Just Use AI?

Include Security analyzes the gap between AI model marketing claims and real-world vulnerability discovery performance in penetration testing. Key limitations persist: false positives are rarely disclosed but add significant review overhead, LLM-based harnesses require substantial tuning beyond simple prompting (one-shot approaches fail on real codebases), and model non-determinism produces inconsistent results across runs. Testing shows high-signal harnesses on small training applications fail to scale to production systems with large code bases, and guardrails from model providers limit frontier model access. Despite AI’s utility for candidate generation, assessments remain expensive because human effort shifts from finding vulnerabilities to validating and triaging AI output—with 65% of practitioners surveyed spending more time deciding when to trust AI results.

Read original article

This post is licensed under CC BY 4.0 by the author.