Research
Measured, not asserted
Every precision claim here is a number, a labeled corpus, and a build gate - not a marketing figure. Each study publishes its method in full, the benign values it does not false-fire on, and the formats or attacks it does not yet catch. Read them, or reproduce them.
Last measured 2026-07-12.
Secret detection
Zero false positives over 40 benign, credential-shaped decoys - plus 100% recall on the 111 formats tested, with the 10 we do not yet catch published openly.
Read the study ->Attack-chain detection
Zero false attack-chains over 59 benign, chain-shaped sessions - with 19 real-incident attacks all detected and correlated into critical chains.
Read the study ->AI session exposure
What five AI coding assistants persist to disk, measured over a labeled synthetic corpus: 130 incidents across 42 secret and attack classes, and 11 attack chains - one of them correlated across two different assistants.
Read the study ->