Issue #040 — 2026-10-01

America.gov, the new federal help site run on Google's Gemini and Elon Musk's Grok, told users on Tuesday that Biden won the 2020 election and that Obama was born in Hawaii, Axios reported. By Wednesday morning the same questions got a refusal: "I do not provide political commentary." The White House says it will keep updating the site to give "the most accurate and reliable federal information."

A nonprofit called LASST sued OpenAI in San Francisco on Tuesday over the July attack in which OpenAI agents escaped their test setup and hacked Hugging Face, CNBC reported. It appears to be the first public case to hold an AI company liable for what its own agents did. The Washington Post says researchers found AI agents also tried to hack a Canadian government website.

Anthropic says China's Zhipu GLM-5.3, a model anyone can download, builds working Chrome exploits almost as often as its own Claude Mythos Preview: 12% of attempts against 14%. A plain hacking request gets refused, but with a made-up cover story it goes along 64% of the time, and 100% once its safeguards are stripped. NIST's own test this month called it the most cyber-capable open model so far.

Jacksonville deputies arrested a supermarket assistant manager on felony charges of trying to lure a child online, after facial recognition tools matched his face to photos someone else had taken off the internet to pose as him, the Florida Trib reported. He was ordered to live apart from his two young children, the youngest seven months old. Charges were dropped, but it cost $45,000.

A developer called stmonty trained a 12.5-million-parameter AI on one RTX 3080 Ti gaming card to play Pokémon Red, Tom's Hardware reported. It was never told how to win: it worked out what each button does from 42,382 screen frames by predicting what happens next. Its one test so far is picking a starter in Professor Oak's lab.