Lakera

Lakera sells security for systems built on models, and its own framing is the useful part: controlling what AI can access is not enough, control what it does. That moves the problem from the familiar list of permissions to behaviour — an agent with legitimate access to a payment system can still be talked into using it wrongly, and no access-control list catches that.

The offering follows: runtime guardrails on application traffic, protection and governance for employee use of AI tools and for enterprise agents, risk-based testing, and a red-teaming service where humans attack the system. The distinction worth holding on to is between testing and runtime defence. Red teaming finds the attacks that existed on the day it ran; guardrails are what keep working on the attacks invented next week, and neither substitutes for the other. The usual caveats apply and matter more than usual here: a guardrail is itself a model making a judgement about input, so it has false negatives, and any vendor's detection rate was measured against a set of attacks the vendor chose.