Falcon uses execution traces to verify secure agentic code
Original title: Secure Agentic Coding through Counterexample-Grounded Feedback
The framework ties security rewards to real program behavior during test-time repair and reinforcement learning.
A cs.CR paper introduces Falcon, a counterexample-grounded framework for verifying secure code from coding agents. It separates attack exploration from behavioral verification and links them with execution traces of the program under test. The authors report stronger held-out functional-security gains than alternative signals in test-time repair, and better held-out results than static-analysis and learned rewards when Falcon is used as an RL reward. Combining Falcon with either proxy reward improved held-out performance further.
Unconfirmed: The held-out gains are results reported by the paper’s authors and are not independently replicated here.
- Falcon separates attack exploration from verification and connects them with execution traces.
- As an RL reward it beat static-analysis and learned rewards on held-out functional security.
- Combining Falcon with either proxy reward further improved held-out performance.
- Authors say policies can optimize a training reward without removing the underlying vulnerability.
From the original article
1 min read at arxiv.orgAs coding models and agent harnesses become more capable, reliable security verification is becoming a central bottleneck. Improving secure code generation requires a signal whose reported progress remains aligned with actual program behavior under optimization; otherwise, an optimizer can improve the signal's verdict without removing the underlying vulnerability.
By Tue Le · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.