Intentest externalizes long-horizon pentest state onto a fact-intent DAG to stop agents looping
LLM agents on long-horizon security tasks lose early critical facts and causal chains to context forgetting and drift into aimless repetitive exploration. Intentest moves that state out of the context window onto a persistent fact-intent directed acyclic graph where verified network states are immutable fact nodes and exploration directions are intent edges bounded by predecessor facts, under a three-layer architecture, and reports substantially fewer invalid transitions on automated web application penetration testing. The immutable-facts-plus-bounded-intents split is a reusable pattern for any long-horizon agent whose failure mode is re-trying what it already ruled out, not just for security work.
Source
↳ Follow the thread