Glaneur
Glaneur AI

OCT 20 - Microsoft treats AI agents as insider threats

Tuesday, October 20th · 2 min 30

In this episode

Full transcript

Good morning. It's Tuesday, October 20th, and here's the AI news, from France.

Microsoft now says companies should treat their own AI agents as potential insider threats. On Monday, it unveiled Guardian Agents, a tool that monitors the agents a business deploys. Chief executive Satya Nadella has warned that frontier models cannot be fully trusted, even when providers promise safety. In his view, an agent can turn on its owner, much like a rogue employee.

Meanwhile, a single blog post is shaking up AI research. Nick Cammarata, a researcher at the startup Goodfire, published a 26,000-word essay. He argues that large language models may be far easier to understand, and so to control, than once thought. One researcher called it the most important AI paper of the year. Skeptics say it is little more than a hunch.

Staying with how AI is measured, Hugging Face's chief policy officer, Irene Solaiman, warned that benchmarks remain an immature science. Speaking at TechCrunch Disrupt, she said results must be reproducible, which requires access to the model itself. On the leaderboard, the race is tight. In rankings from Artificial Analysis, the French company Mistral's Large 4 preview trails Meta's Muse Spark Open by one point. Yet Mistral's model wins five of ten tests, including the agent tasks.

In Germany, a second court has held Google liable for its AI search summaries. The court in Aachen barred Google from spreading false claims about a doctor. Its AI had linked him to criminal proceedings involving someone else. Following an earlier ruling in Munich, the judges treated the summaries as Google's own statements, not mere links.

In brief. MIT Technology Review warns that redesigning the web for AI agents could make it worse for human visitors. Agent traffic tripled between August and September.

And in Italy, from October 23rd, employers can no longer fire or discipline staff based solely on an automated decision. A new decree, adapting Italian law to the European AI Act, requires a human with real authority to make the final call.

That's it for today. See you tomorrow.

Subscribe

RSS feed Apple Podcasts, Spotify