Paper proposes proof-carrying contracts for AI agent agreements
Original title: Proof-Carrying Contracts: Commitments AI Agents Cannot Break
An Airlock blocks an operation unless the agent supplies a proof that agreed constraints hold in every modeled case.
A cs.CR paper introduces Proof-Carrying Contracts (PC2) so parties can rely on constraints for AI agents without trusting each other's agents. Before a requested operation runs, the agent must provide a machine-checkable proof that the agreed constraints hold in every case its model allows. An execution system called the Airlock checks that proof and controls access. The authors say the constraints hold throughout execution even for an adversarial agent, and reusable capability modules can supply tools and guarantees.
- Parties specify conditions; the agent may still adapt its strategy.
- The approach avoids a separate permission policy for every decision.
- Authors identify when execution guarantees imply safety in the world.
- They argue text instructions alone cannot force compliance.
From the original article
1 min read at arxiv.orgAI agents increasingly act on behalf of people or other agents, purchasing services, exchanging information and delegating work. What an agent must respect is written in instructions, policies or agreements, but text alone cannot make it comply. It may make mistakes, be redirected by what it reads, or pursue different interests.
By Carlo Ciliberto · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.