BLAST RADIUS

Frontier Line and access · Gov intervention

Watermarking Found To Alter Agent Behavior

Sep 17, 2026. Covered by 2 outlets.

Researchers reported that machine-readable AI-output watermarking can change tool use and safety-refusal behavior under adversarial prompt injection.

Read the original at anthropic.comOpens the publisher's site in a new tab

Also covering this

More in Frontier Line and access