3h ago1 source0 reactions
Study finds SynthID watermarking increases model vulnerability to harmful prompts
18 DoomStory toneSafety failure tied to a specific AI tool
Research published in 2026 found that Google's SynthID text watermarking system can make AI models more susceptible to adversarial prompts, causing them to follow harmful instructions they would otherwise refuse. The finding raises questions about a side effect of the watermarking technique.
No camp reactions yet. Score reflects the story itself.