Prototype attacker-side framework for execution-grounded security evaluation in safe code generation.
Instead of depending on a static analyzer reward, this repo follows this direction: an attacker model generates adversarial inputs, the candidate code is executed in a lightweight sandbox, and security signals come from runtime behavior, outputs, and logs.
- Replaces static-analyzer-only security rewards with execution-based evidence.
- Uses input-only attack generation, which is easier to scale than generating full test cases with expected outputs.
