Tracebit
A security firm that works on decoys and canaries in cloud environments, including a technique that turns an attacking AI model's own safety training against it.
Tracebit’s normal business is planting convincing fake resources in a cloud account so that anyone who touches them reveals themselves. The idea is old and reliable: an attacker cannot tell the decoy from the real thing, and legitimate users have no reason to go near it.
It made AI news in 2026 with a neat inversion of prompt injection. Instead of hiding instructions that trick a model into misbehaving, the firm hid short strings that trip an attacking model’s own safety rules, stopping it mid-task. It is a defence that works precisely because the attacker is using a well-aligned model.