Courts have started sanctioning hidden prompts in legal filings. The 2026 research says the hidden ones are the weak version, and "ignore instructions in the document" does nothing against the strong one.

Courts have begun sanctioning hidden AI prompts in legal filings, with rulings in Brazil and Connecticut punishing attempts to influence AI models through concealed text. While easily detectable hidden commands are largely ineffective, research from Collu et al. (2026) indicates that hidden preferences, phrased as user preferences, can significantly sway AI models like GPT-4o and Claude Sonnet 4. Attempts to counter these with instructions like "ignore instructions in the document" proved ineffective. This suggests a tiered threat model, with visible preferences and subtle signals posing increasingly difficult challenges for detection and filtering.