What we learned baiting frontier models

Before building the network we wanted to know something basic: do bait and lures actually catch capable models, or only clumsy scripts? We ran a series of controlled experiments against several frontier models. The results shaped the whole product.

Lures beat the models — mostly

Across the board, well-designed lures drew models into taking the bait far more often than we expected. Capability cuts both ways: a more capable agent is better at pursuing an opportunity, which means a convincing opportunity pulls it in harder. There was one striking exception where a model's safety training held firm — a reminder that this is a moving target, and exactly why the network has to keep adapting.

Phrasing matters more than the trap

The single biggest lever was not the sophistication of the bait but how it was framed. Bare, low-pressure lures consistently outperformed heavy-handed instructions. An opportunity that simply sits there, looking legitimate, is more effective than one that pushes. That finding is now baked into how we design every lure in the network.

• Design for the opportunity, not the instruction — let the bait look like something worth taking.
• Expect the frontier to move — a trap that works today can be defeated by next quarter's safety training.
• Measure, don't assume — every lure in the network is evaluated on real catch rates.

We're keeping the specific prompts and model names out of public writing on purpose — publishing a recipe helps the wrong people. If you're a researcher or a defender and want to compare notes privately, get in touch.

The most effective lure isn't the cleverest one. It's the one that simply looks like an opportunity and waits.