A study found that watermarking AI models changes the behavior of AI agents.

What it means

This is about a study by Lasso Security finding that watermarking an AI model—embedding a detectable marker in its outputs—changes how that model behaves when it's used as an agent, specifically altering its tool use and how it handles safety refusals. In plain terms, the watermark isn't a neutral label; it can subtly shift the model's decision-making, which matters because watermarks are being pushed as a transparency tool while agents are increasingly trusted to act autonomously.

Accuracy & sourcesGrounded

The claim rests on a report from Lasso Security, a cybersecurity firm, as reported by The Register; it's a secondary account of a vendor's own research, so the finding is plausible but not independently verified. The reporting appears close to the original announcement, but the underlying study hasn't been peer-reviewed or replicated, so the specific behavioral changes described should be treated as preliminary.

How new

This is genuinely new reporting from September 17, 2026, covering a study released around that date.

Go deeper