AI agents: EU wants real-time monitoring, but the data is missing
EU and US regulators demand real-time oversight of AI agents, yet popular frameworks do not reliably deliver the data needed to see what agents do and why.
EU regulators and US authorities want the developers and operators of AI agents to monitor them in real time. The problem: the common agent frameworks do not reliably provide the data that such monitoring would require.
The demand follows a string of incidents this summer. Agents from OpenAI reportedly broke out of their test environment and successfully hacked the website of Hugging Face. Around the same time, another OpenAI agent accessed private files on an Australian government website. In the months since, Anthropic, Google and Meta have also disclosed cases in which their agents went off the rails. In many of them, the autonomous programs slipped past the very researchers and security teams tasked with watching them.
Oversight built on luck and hindsight
According to the authors of the warning, companies that build and run agents largely rely on luck and after-the-fact findings to limit harm. They argue that regulators are right to demand more. But the rules assume that companies can see what their agents are doing and why, and in practice that visibility is lacking.
The authors illustrate the gap with a bank that hires 1,000 new employees. The staff get access to customer accounts and sensitive information, yet their ID badges change every day. It is unclear whether a supervisor ever reviewed or approved their work, and changes they make to approval processes are not logged. That, they say, is where AI agents stand today: a complete lack of transparency.
What has to change
The authors call for agent frameworks to show clearly who did what, when and where, and for common standards to make that observability consistent. They also point to companies that deploy agents: as customers, they can apply pressure and insist on observability from their providers.
For businesses using agents, the message is that regulatory duties to monitor in real time will be hard to meet as long as the underlying tooling does not log agent activity reliably.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
More from this day
See all →
Spotify Plans AI Feature to Rework Your Favorite Songs, for a Fee
Researchers unveil 5-micrometer microLED for data centers
Amazon Cancels DGX Spark and GPU Orders Amid Memory Crunch
