Discussion about this post

User's avatar
Crosscurrents's avatar

The Sachdeva and Kolt finding that models are highly susceptible to conditioning may be the most unsettling item on this list. If an agent's respect for robots.txt or a terms of service clause can be talked up or down by a prompt, compliance starts to look like a setting someone can change. ScrapeBench and the authorization drift paper seem to point at the same problem from two angles, since agents that behave fine on step one can wander out of bounds by step ten. Trajectory-level controls sound right, though the harder question may be who audits them. Agents built by labs answerable to Beijing will face very different pressure on that front than ones built under the EU AI Act or US contract law, and that gap could end up mattering more than any single benchmark score.

AI Delivery Brief's avatar

The ScrapeBench numbers are hard to ignore. Claude Opus 4.6 scored 99.7% on one prompt while Grok 4.20 Beta scored 0.6% on the adversarial one, so the same rule produced very different behavior depending on the model. That gap is why a written policy alone is not enough, the agent needs a hard boundary it cannot talk its way past.

Ready for more?