AI 智能體的「可驗證」是甚麼意思?
What Does "Verifiable" Mean for an AI Agent?
可驗證的 AI 智能體會留下外部人士可核查、能察覺事後改動的操作記錄。每一步都由執行軟件加上數碼簽署,並以雜湊值連接前一筆記錄;但記錄只能證明發生過甚麼,不能證明操作應否進行,這仍需測試和人工審查。
A verifiable AI agent keeps a record of what it did that someone outside your team can check, and the check shows whether anyone changed it later. That's the bar I'd hold any vendor to, us included. If the only proof is "our admins say the logs are accurate," the agent isn't verifiable.
It's Microsoft Ignite week, and the security track has a session called "When AI acts, security has to answer." I agree with the title. The answer needs evidence that holds up when the person asking doesn't trust you.
Here's how verifiable AI agents do it, in plain terms. Each step the agent takes gets a digital signature from the software that ran it, a bit like a wax seal on a letter. Each signature also includes a fingerprint, called a hash, of the signature before it. Change an old entry and every link after it stops matching.
The strongest objection is that good logs with locked-down access are enough. For debugging, they are. But the people who ask for proof, like an auditor or a customer disputing a charge, sit outside your team, and anyone with admin rights can edit a log.
One honest limit, from Diagrid's plain-English explainer: "the record proves what happened; it doesn't prove it should have." You still need tests and human review for that part. https://www.diagrid.io/learn/cryptographic-attestation-ai-agents
Ask whoever runs your agents who outside the company could check that record today.
來源:DEV Community · dev.to