A new open standard, Agent Evidence Levels (AEL), offers a way to independently verify what an AI agent did, moving beyond simple trust in system logs. It grades how much an outside party can confirm actions and detect omissions without relying on the vendor.
Ever wonder how you can truly trust what an AI agent says it did? It's a tricky question, because most tools just tell you 'we keep tamper-proof logs' and expect you to take their word for it. But what if you don't want to blindly trust the vendor or operator? That's where Agent Evidence Levels (AEL) comes in, a new open standard that helps you verify the actions of AI agents, and what this means for you is greater transparency and peace of mind. AEL isn't just another logging system; it's a way to grade how much an outside party can actually confirm what an AI agent did, and detect if anything was left out. It moves beyond simply trusting a single company's logs. The creator of AEL built it because they faced this very problem with an 'agent firewall' and realized that simply saying 'it blocked the attack' wasn't enough; others needed to see the proof. The AEL standard defines five levels, from AEL-0 to AEL-4, each adding more layers of verifiable evidence: * **AEL-0 (Authentic and Ordered)**: This is the basic level. Records are signed and linked, so you can detect if they've been changed or if something was deleted internally. However, if the keyholder fabricates or truncates the end, you wouldn't know. * **AEL-1 (Gap and Truncation Evident)**: This level adds 'open' and 'close' signatures and 'heartbeats'. This means you can now spot if parts of the record are missing or if the log was cut short. * **AEL-2 (Cross-Domain Omission Evident)**: Here, a second, independent recorder is used with a different verified key. If something important (like a declared event) is missing from one record but present in the other, you'll know. * **AEL-3 (Externally Anchored)**: This level takes verification outside the system entirely. Key parts of the record are registered in an external, public, 'append-only' log. This makes it impossible to present conflicting versions of history without being caught. * **AEL-4 (Counterparty-Confirmed)**: This is the highest level. For specific interactions, the system that received the agent's action actually confirms what it got, even if it was 'nothing'. This doesn't confirm if the action was good or bad, just that it was received. AEL allows you to 'demonstrate' the trustworthiness of an AI agent's record, not just 'assert' it. It’s about building a system where transparency is verifiable, offering peace of mind to anyone relying on AI agent actions. This is a big step towards greater accountability in AI.
