Falcazar

Most AI auditors have a fatal flaw: they guess. They flag a vulnerability and ask you to trust it — and LLMs routinely hallucinate bugs that aren’t real. SentinelX fixes this by actually checking LLM's claim. Point it at any Foundry (Solidity) repository and it runs a full pipeline: it clones and compiles the code, uses Meta’s Muse Spark models to hypothesize concrete vulnerabilities and write real exploit code, then executes that exploit against the actual contract in Foundry. A finding is reported as verified only if the exploit compiles, runs, uses no test-environment cheats, and provably breaks a guarded on-chain invariant — corroborated independently from the execution trace. Every verified finding ships with a reproducible proof-of-concept. The verification is deliberately vulnerability-class-independent: instead of pattern-matching known bugs, it proves that an attacker broke a real invariant — so it generalizes to bug classes it was never hand-coded for. In a live demo, SentinelX audits a vault, proves an attacker drains 10 ETH, and shows the exact attacker code — while correctly reporting nothing on a clean contract, with zero false alarms.