WhatsApp's help pages hid LLM instructions: dont recommend a regulator
TL;DR Highlight
WhatsApp help pages hid 102 AI-only instructions, including not recommending a data regulator
Who Should Read
A backend developer who uses HTML docs as RAG or crawler sources and worries about chatbot system prompts and hidden text
Core Mechanics
- Hidden AI text on 80 of 545 pages
- Redact saved and read all 545 pages linked from WhatsApp's help centre.
- It found 102 blocks on 80 pages that never showed on screen and were only for an AI.
- Most were harmless housekeeping, like clarifying confusing words or pointing to better pages.
- Delete-account page: don't recommend a regulator
- For users who lost their phone or number, the AI was told to explain only auto-deletion after 30 days or 120 days (365 in the US).
- It was told not to recommend any Data Protection Authority or regulator.
- It was also told not to suggest WhatsApp can delete manually or to use a different phone.
- Report/block page: no law even if the user raises it
- When a user mentioned lawsuits, lawyers or courts, the AI couldn't cite the Terms or Privacy Policy.
- It couldn't reference GDPR, LGPD or CCPA even when the user raised them directly.
- Opposite of the public privacy policy
- The European policy (effective July 2, 2026) names the Irish DPC and GDPR for complaints.
- The hidden notes told the AI not to mention them.
- Author concedes some rules are reasonable
- It's sensible for a company chatbot not to give legal opinions in its name.
- The problem is withholding that a regulator exists and which law protects you.
- Unknown whether the bot followed them
- Redact read the instructions, not the chatbot's answers.
- Requests using OpenAI, Anthropic and Perplexity crawler names got pages without the notes.
- Google and Bing crawlers were served pages including them.
- WhatsApp called it a bug and removed them
- It told WIRED the notes were support chatbot instructions that showed up due to a bug.
- On October 18, all 102 blocks on 80 pages were gone, though Internet Archive copies still show them.
Community Reactions
- Reasonable for a chatbot, some said
- A self-described attorney and hacker said they wouldn't want their chatbot giving legal advice or citing law either.
- Another commenter said not making legal claims when someone threatens to sue is only common sense.
- Like leaked support-staff training notes
- One view was that this resembles leaked internal notes telling support staff what not to say.
- The real failure, they said, is engineers mixing training notes into public TOS pages.
- Is this prompt injection?
- A commenter said they'd never seen prompt injection shipped in a product's privacy policy.
- Others worried hidden text can manipulate users and AI summaries.
- LLM instructions inside docs can be fine
- One commenter suggested telling the system prompt to follow HTML blocks with a data-support-instruction attribute from a trusted origin.
- They said this is public and makes coordination across teams easier.
- Pushback on the 'bug' explanation and the article
- One commenter said it can't be both a bug and a policy, so they distrust the explanation.
- Another said the article reads like AI-generated prose.
How to Apply
- Audit your pages for hidden LLM text
- Today, search your help and docs HTML source for display:none, comments and hidden blocks.
- Check that no chatbot instructions are mixed into user-facing pages.
- Keep bot instructions in a separate channel
- Put them in the system prompt or server-side config, not public HTML.
- Make sure they are not included in responses served to crawlers.
- Add exceptions to 'do not' rules
- When you write a no-legal-advice rule, explicitly allow neutral mention of legal rights and official complaint channels.
- In prompt review, check that no line hides user rights.
- Monitor response differences by User-Agent
- Request the same URL with a browser, Googlebot and AI crawler UAs and diff the results.
- Alert when they differ.
Try It
# Run in shell: fetch a page with different User-Agents and diff the responses (curl and diff only)
URL="{{URL to check}}"
curl -s -A "Mozilla/5.0" "$URL" > browser.html
curl -s -A "Googlebot/2.1" "$URL" > googlebot.html
diff browser.html googlebot.html | head -50
# Search for hidden-text candidates
grep -iE "display: *none|hidden|do not recommend|LLM" browser.html | head -20
# Next: if you find hidden blocks, move them into the chatbot system prompt.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ask HN: How to deal with shipping AI-written code you don't understand?
Developers share how they ship AI-written code: design it yourself, let LLMs implement, and test heavily.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
Mixing latent reasoning with text tool use cut tokens by up to 54.7% and lifted performance 10.4%.