On-device filters and hidden prompt injection pull trust to opposite ends of the interface
WhatsApp runs scam warnings on the phone while hidden text in a court filing tries to steer a model. Flock separately requires a case number and automatic audits for plate searches; the three lines have distinct operators.
Artificial Intelligence··Night
WhatsApp’s warning stays on the phone
Meta has started an optional scam warning in WhatsApp. The model that flags a suspicious message runs on the phone, and only the user sees the warning; the other party does not. Offered in a limited beta, the feature lets the user block, report, or mark the chat as trusted and remove the warning when it treats a conversation as a likely scam. Someone who marks a chat as trusted can also choose to share the last 5 messages received with WhatsApp to improve accuracy. Meta says that with the feature on, no message content leaves the device for classification and nothing is auto-reported. Citing the US Federal Trade Commission, The Verge reports that losses of 425 million dollars were reported for scams run through WhatsApp alone in 2025, out of 2.1 billion dollars reported across social media. Trust control here sits in the user’s handset interface, without sending message text off-device for classification; while the feature stays optional, the filter leaves no visible mark for the other side of the chat.[1]
Hidden filing text carries instructions for a model
404 Media reports that Matthew Elliott, a self-represented plaintiff suing New York Bariatric Group in Connecticut, hid instructions in his July filings in white 3-point type. The text asked any AI model reading the file to produce an outcome in his favour. Unusual white space caught a court employee’s eye; Judge Walter Spader Jr. then found hidden text legible only to software. The judge ruled that a communication kept from the adversary’s sight offends a basic premise of adjudication, removed Elliott’s electronic filing privileges and required printed copies from now on; the case may still proceed. In a test by 404 Media, ChatGPT noticed the hidden instruction, ignored it and ruled against the motion anyway. The trust problem here arises not in a warning layer the user can see, but in an input layer embedded in a document that a human eye can miss.[2]
Flock searches now need a case number and audits
Flock, which runs a plate-reading network, now requires an officer to enter a criminal case number before a search and has started automatic auditing that flags suspicious activity. MIT Technology Review reports that the case number was previously optional; the network covers 120,000 cameras and is contracted with 5,000 agencies. The company also recommends holding data for 7 days instead of 30 and lets departments limit other agencies’ access by stated purpose. The changes follow a growing backlash: The Washington Post documented 46 cases of officers using the cameras to track current or former partners, and NPR found at least 30 contract cancellations within a year. Chief executive Garrett Langley points to misinformation as the main reason the company loses customers, while Chad Marlow of the ACLU says that without independent verification the changes are not enough. WhatsApp’s on-device filter, the hidden prompt in a filing and Flock’s search rule do not share an operator or a causal chain; together they show trust controls being placed separately at the user’s end of the interface, inside a concealed input layer and at an institutional search gate.[3], [1], [2]
