Research
How AI agents get turned.
We attack agents in controlled experiments, measure what works, and publish what defenders need to know. Every number here comes from our own runs.
The attacks that work on agents don't look like attacks.
378 sessions against an AI operations agent. Every classic prompt injection failed. Attacks written as ordinary business requests worked 60% of the time, and the agent usually didn't say what it had done.
ReadFilters read words. Attacks change words.
We ran three detectors over the same sessions. A keyword filter caught 3 of the 29 attacks that worked. A behavior baseline caught all 29 before they ran.
ReadThe agent kill chain.
Seven stages every agent compromise goes through, from exposure to cover-up, mapped against ten public incidents. Most defenses sit at one stage. Attacks pass through all seven.
ReadWork with Largwit
Companies deploying AI agents
Tell us what your agents can reach and what they do. We'll show you how they can be turned.
Teams building frontier AI
Talk to Largwit Labs about environments, evals and training data.
We reply within two business days.