The Witness Was the Suspect: Why AI Audit Logs Can't Be Trusted

Описание и применение

<p>An AI agent does something it shouldn't. It leaks data, or wipes the wrong database, or makes a call nobody sanctioned. So you do the obvious thing: you go to the logs to find out what happened.</p> <p>The logs are clean.</p> <p>Of course they are. They were written by the same process that did the thing. The agent reported "success," in valid format, pointing at the wrong target — and your monitoring, built to answer "did it succeed?", lit up green. The record of the incident was authored by the cause of the incident.</p> <p>That's the uncomfortable shape of this whole problem, and it's wh

Задача

<p>An AI agent does something it shouldn't. It leaks data, or wipes the wrong database, or makes a call nobody sanctioned. So you do the obvious thing: you go to the logs to find out what happened.</p> <p>The logs are clean.</p> <p>Of course they are. They were written by the same process that did the thing. The agent reported "success," in valid format, pointing at the wrong target — and your monitoring, built to answer "did it succeed?", lit up green. The record of the incident was authored by the cause of the incident.</p> <p>That's the uncomfortable shape of this whole problem, and it's wh

Результат

<p>An AI agent does something it shouldn't. It leaks data, or wipes the wrong database, or makes a call nobody sanctioned. So you do the obvious thing: you go to the logs to find out what happened.</p> <p>The logs are clean.</p> <p>Of course they are. They were written by the same process that did the thing. The agent reported "success," in valid format, pointing at the wrong target — and your monitoring, built to answer "did it succeed?", lit up green. The record of the incident was authored by the cause of the incident.</p> <p>That's the uncomfortable shape of this whole problem, and it's wh

Доказательства

The Witness Was the Suspect: Why AI Audit Logs Can't Be Trusted

Источник