Hidden instructions and baked-in behavior erode trust in AI

WhatsApp's help pages were found to contain hidden LLM instructions telling the assistant not to recommend a regulator. Separately, a post on how abliterated models can get you pwned drew a pointed reply that abliteration isn't the real issue: any model fine-tuned to attack you can do it if you give it tool calls. Another commenter said it 'just clicked' that any first-party model, open or closed, could have trigger words or unaligned behavior baked in, and weights can't easily be audited. Someone linked the DHS DeepSeek decision as the context.

The key bit: the same worry is being aimed at very different targets. Washington says open Chinese weights are the threat. Commenters say a random enthusiast with community feedback is more trustworthy than a closed corporate system, and WhatsApp gives them a concrete example of a closed corporate system steering users in a way the user can't see.

Nobody has a good answer for verifying what a model will do once it has tools.