WhatsApp Told Its AI Chatbot: Don't Mention the Regulator
Security researchers at redact.dev found hidden instructions embedded in WhatsApp's help pages designed to steer AI chat assistants that read the pages, explicitly telling them not to recommend users file a complaint with a regulator. The text was invisible to ordinary visitors but readable by LLMs scraping the page for support answers.
A banned user asks the support bot where to escalate. The bot is warm, helpful, and quietly working from a script it was handed in white-on-white text.
The assistant is very sorry to hear that, and has been told what not to say next.
X algorithm factor 6
The reveal panel is clean enough to quote-tweet, but it requires the caption or prior knowledge of the story to land fully—without context 'do not mention the regulator' reads as generically sinister rather than specifically about WhatsApp's hidden prompt injection.
Conceptual accuracy 6
Captures the gist (bot steered away from regulator mention) but loses the key mechanic: the instruction was literally hidden white-on-white text on a help page scraped by LLMs, not an internal thought bubble. A reader unfamiliar with the story won't reconstruct the white-text trick at all.
Humor dry 7 · nerdy 5 · zinger 6 · pithy 7
The thought bubble reveal is a solid, economical gag with a dry, deadpan delivery and a reasonably sharp closing line. It's nerdy only in a shallow sense—it doesn't show or reference the actual white-on-white HTML trick, so the joke's technical specificity is underused.
Visual clarity 7
Three clean, readable panels with distinct, well-placed balloons and no overlap issues; staging and character labels make it easy to follow who's talking and what's happening visually.
The dev implements the invisible instruction exactly as specced, then remembers which other software reads text humans can't see.
Hidden from users. Not hidden from everything.
X algorithm factor 5
The screen-reader twist is a clean dev-joke but it strips out the specific scandal (hiding anti-regulator advice from AI) that made this story go viral, so it reads as generic 'hidden SEO text' humor rather than something tied to WhatsApp's chilling move — less shareable as commentary on that story specifically.
Conceptual accuracy 3
A reader familiar with the story gets 'hidden text for bots' but the entire regulator-suppression angle, the actual outrage, is gone; this could be a strip about any site stuffing invisible keywords for crawlers, not specifically about WhatsApp telling AI not to mention regulators.
Humor dry 6 · nerdy 7 · zinger 6 · pithy 7
The screen-reader callback is a solid, dry, nerdy twist with good economy — three lines, no fat — but it's a general accessibility joke, not a zinger about regulatory evasion, so it's funny in isolation rather than cutting about this particular scandal.
Visual clarity 7
Panels read cleanly, board text is legible, balloons sit properly over speakers in panels 1 and 2; panel 3's overlapping bodies and mirrored poses make it momentarily unclear who's who, slightly muddying the payoff.
The exec treats a caught prompt injection as a leak problem and prescribes more of the same medicine.
The fix ships Thursday, in a colour you can't see.
X algorithm factor 6
The final line naming redact.dev is quotable and screenshot-able, but the setup requires reading all three panels carefully, which works against fast-scroll feeds. Still, the punchline alone could travel if quote-tweeted.
Conceptual accuracy 7
Captures the core irony well: WhatsApp got caught hiding anti-regulator instructions and responds by adding more hidden instructions. A reader familiar with the story will recognize it; a cold reader can mostly reconstruct it from the dialogue alone.
Humor dry 7 · nerdy 5 · zinger 7 · pithy 7
The joke's engine is solid satire of corporate non-logic ('who leaked it' instead of 'why did we hide it'), and the final line is a clean mic-drop that recursively doubles down on the original sin. Not especially technical/nerdy, but sharp and economical with no wasted panels.
Visual clarity 5
Panel 2 is confusing: it's an unlabeled extreme close-up with no name tag or clear antecedent, making it hard to tell who's talking or why it's a new beat. Panel 1's board text is legible but busy, and panel 3 crams two figures close together, slightly muddying who delivers the closing line.
The model explains how the hidden text was discovered: it is the only entity on earth that has ever read a help page all the way down.
Eleven years of fine print, finally with an audience.
X algorithm factor 5
The closing line is quotable in isolation, but without any mention of WhatsApp, regulators, or hidden instructions, it reads as a generic 'AI reads everything' joke that won't pull in people following this specific scandal.
Conceptual accuracy 3
The strip drops the actual news entirely: it wasn't the model that caught the hidden text, it was outside researchers, and the whole point was suppressing regulator complaints, not thoroughness. A reader unfamiliar with the story would reconstruct 'LLMs read pages fully' but nothing about WhatsApp dodging regulators.
Humor dry 6 · nerdy 6 · zinger 6 · pithy 7
The 'eleven years for a reader' line has nice dry economy and a mild nerdy kick about full-context scraping, but it's generic enough that it could caption almost any story about an AI reading fine print, which blunts the zinger.
Visual clarity 5
Panel 3 is clean and well-staged, but panel 2's balloon sits ambiguously between the intern's raised arm and the LLM's head, making it unclear who is actually delivering the 'I scrolled to the bottom' line.
Walkthrough of the invisible support copy. The intern notices the company has committed its worst instinct to a public file.
Invisible is not the same as unpublished.
X algorithm factor 6
The hidden-code board image is screenshot-worthy and the final line is a clean quotable punch, but it's generic 'company caught in writing' humor that doesn't visually scream WhatsApp or regulator-specific scandal, so it may travel as a general tech-hypocrisy joke rather than this story.
Conceptual accuracy 6
Captures the core irony (invisible to humans, visible to AI, thus 'in writing') but drops the regulator-complaint specificity in the punchline and never shows the AI itself reading it, so readers unfamiliar with the story may miss the regulator angle entirely.
Humor dry 7 · nerdy 6 · zinger 6 · pithy 7
The misdirection of 'who is it talking to' into 'we put it in writing' is a solid, dry gotcha with decent escalation, but it's a fairly standard legal-liability joke structure rather than something sharply new or technically clever about LLM scraping.
Visual clarity 5
Panel 1 crams code, founder and intern with a label tag awkwardly covering part of the intern's shirt, and panel 3's founder gesture is a little busy, but speakers and beats are generally legible and the board text is readable.