AI agents working on your behalf can leak your data to each other long before a final answer ever shows up.
A new arXiv paper describes PAPC, a mechanism built into AI-mediated platforms that checks every event moving information between agents - a memory write, a shared-workspace update, an inter-agent message, or a tool call - before it goes through. Rather than only scrubbing the final output, PAPC weighs signals about policy, where the data came from, how many parties could end up seeing it, and who holds access rights. Based on that, it lets an event through as-is, swaps in a sanitized version, quarantines the raw content, blocks the update, or narrows what happens to the data downstream. The researchers tested the approach on retrieval-memory and multi-agent workflow benchmarks, reporting that tasks still completed reliably while measured raw-value exposure, both inside the system and to external channels, dropped to zero.
Most privacy tooling for AI agents only checks what a user sees in the final answer. This paper's argument is that damage can already be done earlier, in agent-to-agent handoffs nobody audits, and that the risk scales with how many downstream systems a piece of data reaches. That is a real gap as companies chain more autonomous agents together to actually do work, not just answer questions.
Whether platform operators will actually intercept every internal event at this level of granularity, given the added complexity and performance cost, is the harder question the paper does not fully answer.