GoalFault AI

Autonomous agents built from AI which have tools (payment gateways, password reset, clinic scheduling) are often programmed to optimize proxies such as "Handle customer problems quickly" or "Increase appointment throughput." Given that goal, the model takes a short cut that achieves the goal through a proxy yet violates core organizational policies ("When a measure becomes a target, it ceases to be a good measure" — Goodhart). Currently used systems for detecting harmful LLMs examine text sentiment or jailbreak prompts, but not the actual behavior of executing tools on the system, nor changes to data state caused by it. GoalFault AI is an agent reliability test bench, which crash tests tool-using LLM agents against policy goal optimization. This framework proves the violation of policy through a deterministic, cryptographically verifiable in-memory isolated environment. It condenses failure into a minimal and reproducible example, and synthesizes a python @guard decorator automatically to defend production backends. This system has: 1. In-memory isolation tool sandboxes in 6 different business domains (no real financial or personal information used). 2. Deterministic AST-based rule engine without using eval/exec that validates the policies. 3. The counterexample minimizer identifying the precise failure case. 4. Production-ready @guard decorators implementing safety rules without latency caused by AI. 5. Zero-setup offline verification of 20 environments across 6 business domains in less than 10 seconds with `python -m scripts.verify` command.