vladimir_gor

Really interesting direction. What resonated with me is that you're treating agent security as an information-flow problem rather than a prompt-classification problem. It was not so obvious to me.

A key question I agree isn't just "is this tool call allowed?", but "given everything the agent has read so far, is this information now allowed to flow to this destination?" That feels like a much more fundamental abstraction.

The part I'm particularly curious about is how this will work with policy authoring at scale. What would be the main adoption challenge?

I'm a big fan of benchmarks and now finally we have one to evaluate models on physical tasks. Will be interesting to see how fast this gap will narrow.