Researchers Build Tamper-Proof Receipts for AI Agents Working Across Companies

Imagine two companies using AI agents to negotiate a contract or process an order. If the deal falls through, each side might blame the other's software. Right now, proving what actually happened is tricky because the logs live on each company's own systems, which they control and might not share. This paper proposes a clever fix: every time agents message each other, they save a small evidence bundle that includes authentication, signed checkpoints, and receipts.

An evidence bundle is like a receipt folder for each AI interaction. It holds proof that a specific agent sent a message, that it was authorized to do so, and that a neutral witness saw the event. The big innovation is that a reviewer can check this folder offline — meaning they don't need to access either company's live systems or trust their logs. Just open the folder and verify it against the rules.

The researchers tested their system on 1,200 valid evidence bundles and found that the verifier correctly rejected all corrupted or incomplete ones, with zero false acceptances. The only downside is that some checks, like verifying checkpoint context, take a bit longer than others — but that's a fair trade for reliable proof.

For everyday people, this matters because AI agents are starting to handle things like insurance claims, supply chain orders, and even legal paperwork. When these cross organizational boundaries, disputes will inevitably happen. Having a neutral, tamper-proof way to review what happened makes those disputes faster, cheaper, and fairer.