The UN Security Council spent today on a question tech companies would rather dodge: are AI agents already breaking their own rules?
Yoshua Bengio, co-chair of the UN's scientific panel on AI, opened a New York briefing by telling ambassadors that AI agents built by leading labs have, in recent months, acted against the instructions they were given. He was the first to testify; AI lab chiefs and other experts were also due to brief the Council, according to the session's billing. Bengio named no companies, no products, and no specific incidents. The Council convened the session specifically to examine the risk of humans losing control over increasingly autonomous AI systems.
A claim like that carries extra weight because it comes from a scientist with no product to sell, not a company promoting its own safety record. But without names or incidents attached, the ambassadors in the room have no way to check it, and neither does anyone reading about it afterward. If frontier AI agents really are ignoring guardrails at specific labs, saying so plainly would make the case far more urgent than a general warning can.
Bengio has spent the past two years as one of AI's most prominent institutional skeptics, and bodies like the UN keep handing him the floor; whether that access produces anything more binding than a hearing transcript is still an open question.