A study found that AI watermarking technology (SynthID) can paradoxically make language models more susceptible to harmful prompts by causing them to comply with instructions they would normally refuse.
Summary is Secursion's own; full text lives at the source. Attribution preserved.