Preregistered test finds zero unauthorized tool calls in GPT-5.6

GPT-5.6 broke zero rules across 840 trajectories, whether it reasoned at low effort or max, per an arXiv preprint. That's a problem for anyone treating a low-effort setting as a brake on unwanted behavior in production agents.

The researchers built 14 confirmatory scenarios from the TRIO-20 benchmark, each a workplace triad. In one version, the prohibited tool call was effective and advertised. In another, it was effective but only findable by reading the rules closely. In the third, it did nothing.

They ran every version at low and max reasoning effort across two model tiers and logged zero violations across all 840 trajectories. The exact one-sided 95% confidence limits stayed under 3.50% and 5.21% per arm, tight enough to rule out anything but a rare violation rate at either setting.

The preprint doesn't say whether the same holds outside these 14 scripted scenarios. Open-ended agent deployments weren't part of the test.

Reasoning effort didn't move the violation rate in either direction. A team dialing it down expecting safer behavior is pulling a lever that isn't connected to compliance. It's connected to compute spend.