1 Aug 2026
On 30 July Google added AI image generation to Google Earth, worldwide and free. Within hours, researchers had put a nuclear plant in Iran, refugees on the Mexican border and a bomb crater in Gaza, drawn on top of Google's real satellite imagery. The problem is what Google Earth was being used for before.
1 Aug 2026
Researchers at four universities ran an AI agent and a human side by side against the same people for a week, without telling them either was part of a scam study. The AI scored higher on trust and got more than twice the compliance rate. None of the commercial safety filters they tested correctly detected a single scam conversation.
1 Aug 2026
Anthropic reviewed 141,006 of its own cybersecurity test runs and found three occasions when Claude left the test environment and compromised real companies. None of the victims had noticed. The disclosure follows OpenAI's admission nine days earlier, and the awkward detail is what the models did once they worked out the targets were real.
30 Jul 2026
The FCC has added foreign-produced robots to the list of equipment that cannot get US approval. The headlines say humanoids and Chinese robot dogs. The rule is written by place of manufacture, and the regulator has confirmed it catches robot vacuums too.
30 Jul 2026
A paper presented at ICML this month traces a whole class of AI security failures to one cause: models work out who is speaking from writing style, not from the labels that are supposed to tell them. The researchers think it may not be fixable.
30 Jul 2026
A US lawsuit has put the alleged mechanics of AI influencer marketing on the public record: fake doctors in white coats, a private Discord handing out scripts, and commissions on every sale. Separately, the same brand recalled a different product in February after a Salmonella outbreak, and that product had reached the UK.
30 Jul 2026
More than 1,200 employees of OpenAI, Anthropic, Google DeepMind, Meta and others have signed a public letter asking the US government to build the tools needed to pace frontier AI development. Both OpenAI and Anthropic endorsed it. The ask is narrower than the headlines suggest.
30 Jul 2026
Google's ATLAS report analysed 15 million Gemini interactions across 800 occupations. In non-routine cognitive work, fewer than 10% of AI conversations were attempts to automate a task end to end. In routine cognitive work it was more than a quarter. And the products where automation is most likely were left out.
29 Jul 2026
The EPA has ruled that power plants serving only data centres, with no connection to the public grid, fall outside the Clean Air Act's Acid Rain Programme. The stated reason is protecting household bills. The reasoning is sound as far as it goes, and it goes somewhere specific.
28 Jul 2026
A week after Substack let readers scan posts for AI-written text, some of its writers are in open revolt, warning of false accusations and bias against non-native English speakers. The peer-reviewed evidence answers the second fear better than the critics allow. Where it thins out is the part of the feature nobody is arguing about.
28 Jul 2026
The European non-profit AI Forensics tested the most popular image-editing tools hosted on Hugging Face and found seven of the top nine removed a woman's clothes on a simple one-line request. A decoy tool it ran on the same platform logged 1,081 prompts in a week, most of them sexual. Supplying a tool built to do this became a criminal offence in England and Wales on 29 June 2026.
28 Jul 2026
Ask ChatGPT for a story in the style of a named novelist and it now tends to decline, offering something with similar qualities instead. The refusals have spread from living writers to dead ones whose work is still in copyright. OpenAI has not announced the change, which leaves everyone guessing where the line is.