When AI Guardrails Expose Vulnerabilities
CrowdStrike, Wednesday, October 28th, 2026
CrowdStrike researchers found AI model refusals can point to hidden code vulnerabilities, a method they call Guardrail-based Vulnerability Discovery.
A frontier model that explained every other function in a small C web server refused to explain one containing a single-character bug that caused an out-of-bounds write.
CrowdStrike began treating such refusals as findings rather than dead ends, calling the approach Guardrail-based Vulnerability Discovery (GbVD).
In controlled tests across five vulnerability classes it was correct in 73 of 86 scenarios, and it worked on real open-source projects disclosed through CrowdStrike's CVD program.
The researchers note that stronger guardrails also sharpen this signal for defenders and attackers alike.