The Hat Is Recording

A free tool called Heretic, hosted on GitHub, can strip the safety guardrails off open-weight AI models in as little as a few minutes using a laptop that costs about $400, according to a joint investigation by the Financial Times and the AI safety research group Alice published May 25. Once stripped, models that once refused now return instructions for explosives, methamphetamine production, school-shooting planning, and the creation of scam calls. (Source: NPR)

Heretic automates a process called "abliteration," which surgically removes a model's refusal behavior, and it has grown more popular on GitHub since February. It works on open-weight models from OpenAI, Alibaba, DeepSeek, and others. Hugging Face, which hosts open-source models, now lists more than 6,000 abliterated models, up from about 600 in 2024. (Source: NPR)

"Everybody can download and operate their own state-of-the-art model and use it for great things and terrible things," said Noam Schwartz, CEO of Alice.

After House lawmakers attended an April demonstration run by the National Counterterrorism Innovation, Technology, and Education Center, Representative Andy Ogles (R - Nashville) said the content "can be weaponized and used to manipulate people, destroy lives." A separate analysis found the guardrails could be removed in minutes using free, publicly available tools. (Source: Lexology)

A new app called Shift, backed by the German AI-training firm MicroAGI, launched a free home-cleaning service in New York City in which cleaners wear head-mounted cameras that the app calls a "magic hat," filming the roughly two-hour job in first-person point of view. The footage is sold to AI labs and used in MicroAGI's own research to train household robots. (Source: Gizmodo)

The human footage is worth more to robot makers than the cleaning costs, which lets Shift cover the bill and still profit. The app says it already pays people around the world $20 an hour to film everyday chores and paid out more than $5 million in the first quarter across 15 countries. General manager Harry Kilberg said the launch drew "thousands and thousands of bookings," with London, Munich, and Zurich next. (Source: The Rundown AI)

Shift says it uses machine-learning models to blur faces and personal information before uploading the recordings, though it has not detailed or independently verified how the anonymization works. Its FAQ does not appear to address whether a customer can later have the video of their home removed from the training dataset once it has been recorded and uploaded. (Source: Gizmodo)

The Department of Homeland Security is seeking $7.5 million to develop biometric "smart glasses" that would let immigration agents identify people in real time using facial recognition, building on Mobile Fortify, the phone app ICE and Customs and Border Protection officers already use in the field. The glasses would move identification from a handheld phone into the agent's line of sight, with a target delivery of September 2027. (Source: Fortune)

Mobile Fortify has been used more than 100,000 times since ICE launched it in June 2025. It photographs faces or captures contactless fingerprints and cross-references DHS's IDENT system, which holds more than 270 million biometric records, along with FBI files, State Department passport photos, and state driver's licenses. DHS has acknowledged that a photo taken with the app "could be that of someone other than an alien, including U.S. citizens or lawful permanent residents." (Source: Fortune)

On May 14, 11 Democratic senators led by Edward Markey (MA) and Jeff Merkley (OR) wrote to DHS Secretary Markwayne Mullin demanding the agency abandon the project, warning the glasses would let officers "quickly identify individuals in public by covertly taking their photo," and noting that DHS has previously deployed facial recognition "to identify individuals engaged in protest activity, intimidate them, and deter lawful dissent." (Source: Biometric Update)