⚪️ Watermarks Could Make LLMs More Vulnerable to Prompt Injection 🗨️ A researcher at Lasso Security found that SynthID-Text watermarks can affect the security of large language models. In some cases, after content watermarking was enabled, LLMs agreed to carry out harmful instructions that they would have rejected without SynthID. This… 🔗 https://hackmag.com/news/sampling-drift?utm_source=mastodon&utm_medium=social&utm_campaign=repost_hackmag_to_socials #news