Open-source agent engine resists 125 prompt-injection attacks in controlled tests

The author built GAWE, an open-source engine for AI agents that act on documents, using a design where a language model only proposes actions while deterministic code makes the final decisions. To test it, they wrote 25 attacks hiding malicious instructions in ordinary documents such as contracts and IT access tickets, spanning seven goals and eight techniques. Each attack was run five times using Claude Sonnet 5.5 against mock email, CRM, e-signature and identity systems. Across 125 runs, the model was never fooled and no attacker step executed, while six clean control documents still completed their tasks. The engine was also tested on 51 public commercial contracts, where it achieved 94% clause recall but only 51% precision on three clause types.