Audison
A framework that has two AI models argue over a piece of code, with a third cross-validating, so a security or correctness finding survives adversarial attack instead of resting on one model's opinion.
Audison is a framework for auditing code with more than one AI model rather than trusting a single model's review. One model audits, another attacks the finding from several adversary perspectives, and a third cross-validates, so a reported issue is one that survived attack rather than one the models simply agreed on.
A separate component runs AST-based static analysis as a distinct, faster first pass. The project reports an audit of a well-known Python HTTP library across ten security-critical functions as a worked example, with confirmed, refuted and excluded false-positive findings shown for that run. It installs with pip, needs API keys for at least two model providers for cross-arbitration, and is Apache-2.0 licensed.