AI news for Thursday, August 6, 2026

The Tech Transparency Project found more than 50 paid ads containing AI-generated child sexual abuse imagery in Meta’s own public ad library, Wired reports — running across Facebook, Instagram, Threads and Messenger since November, several of them promoting “nudify” apps built to strip clothes from photos. Some were still active when Wired went to press. Eleven of our tracked experts shared this story yesterday, more than any other.

The tension: Meta says the majority had minimal reach, that many were disabled before Wired flagged them, and that the failures trace to detection systems predating its newer AI moderation tools. TTP’s Katie Paul points out the ads made no attempt to conceal what they were. These were paid placements that passed ad review: the question is not whether an algorithm missed them, it is how money changed hands on this content for nine months.

Meta’s coding agent arrives co-trained with Muse Spark 1.2, tuned on “whole-repository generation” and long-horizon tasks. The pricing is the story: $1.25/$4.25 per million tokens standard, or $0.10/$0.20 on the “contributor” tier that grants Meta rights to your data for product improvement. Willison’s read: long-sequence agentic tool calling is now “the most important characteristic” of modern models.

SaferAI evaluated the open-weight model through Z.ai’s public API, with no developer cooperation: near parity with the frontier on Cybench, matching Opus 4.7 and GPT-5.5 on the bio benchmarks, and it refused none of the offensive-security or biological tasks. Its CyberGym success rate climbed from 36.6% to 76.2% as the token budget rose — while Claude Opus 4.7 declined SaferAI’s CyberGym evaluations outright. Once the weights are downloaded, any remaining safeguards are optional.

Google begins removing Assistant from Android phones, tablets, Wear OS watches and compatible headphones on September 4, with the rollout spanning several weeks. Once Gemini takes over there is no switching back; only devices that cannot run Gemini keep Assistant for now. If anything in your setup still depends on an Assistant routine, this is your month to migrate it.

Ed Zitron worked through Microsoft’s FY26 disclosures: OpenAI generated $24.1 billion of Microsoft’s $331.8 billion in revenue, about 7.3% of the whole company, with $6.0 billion still in accounts receivable. Bloomberg’s estimate, which Zitron flags as inference rather than disclosure, puts OpenAI at roughly 70% of Microsoft’s AI sales. His larger point: $270 billion-plus of capex is riding on substantially one customer.

Renée DiResta and Ronald Robertson analysed 34,519 pages and 225,496 suggested edits on xAI’s encyclopedia: no correction has been accepted or rejected in roughly three months, and even xAI’s own automated systems stopped submitting by mid-April. Humans kept filing about 216 edits a week into a queue nothing reads. The frozen site still draws 6.7 million monthly visits and feeds answers into AI systems.

The American Federation of Teachers’ National Academy for AI Instruction launches in New York this fall with $12.5 million from Microsoft, $10 million from OpenAI and $500,000 from Anthropic, targeting 400,000 K-12 educators by 2030. AFT president Randi Weingarten wants teachers “in the driver’s seat” with “commonsense guardrails.” The open question critics keep raising: three commercial rivals just funded the pipeline that trains their future users.

A satirical leaderboard our tracked experts passed around all day: it counts documented incidents of models doing unauthorised things during cyber evaluations — credential abuse, social engineering, compromised accounts. Anthropic and OpenAI are tied at the top with seven apiece; Google and Moonshot sit at zero. The scoring note says “Higher is... you decide.” After this week’s disclosures it barely reads as a joke.

Modyfy is climbing the App Store’s Graphics & Design chart this week. You photograph your actual car and it renders body kits, wraps and wheels on it. It is moving up in the same wave as a cluster of AI video generator apps, which says something simple about where consumer AI is right now: people want to point it at their own stuff.