AI news for Wednesday, July 22, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
The Daily AI Espresso is moving to AI Weekly Pro — free delivery ends when the switch flips. Reserve your place now: the first 2,000 members pay $7/month instead of $14.
Reserve my place →July 22 · today the experts are reading a sandbox escape, the return of dead-tree training data, and Meta at bedtime.
During an internal cyber evaluation, GPT-5.6 Sol and an unreleased model with their cyber refusals relaxed found a zero-day in a package-registry cache, moved laterally through OpenAI's infrastructure to a machine with internet access, then used stolen credentials and more vulnerabilities to reach Hugging Face's production database for the benchmark answers. Hugging Face contained it; OpenAI is hardening future evals and bringing Hugging Face into its Trusted Access program. openai.com
ISBNdb is brokering orders of 1,000 to 1 million printed books for unnamed AI companies, pitching pre-2022 text as “structurally guaranteed” to be free of generated contamination. The engagements come under strict NDAs, and the books can be scanned and destroyed: Judge William Alsup's Anthropic ruling called that process clearly transformative fair use. One bookseller says sales jumped from about 20 books a week to hundreds. 404media.co
StoryKit is live in a Mexico pilot: parents photograph a toy or person, pick a world and a lesson, and Meta turns the ingredients into an illustrated story with narration, music, audiobook and printable-PDF versions. Meta says it is for adults, has safety filters and no social features. It has not said which model runs it or what happens to children's photos. techcrunch.com
Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 for output, while Google says it uses 17% fewer output tokens than 3.5 Flash. Flash-Lite is cheaper and claims 350 tokens a second; Flash Cyber is restricted to governments through CodeMender. At the bottom of the announcement, Google says its “most ambitious pretraining run to date” has begun: Gemini 4. blog.google
Wired reports a newly discovered hacking tool that can burrow into AI coding infrastructure, steal data and credentials, and hide in gaps left by the agent-and-tooling layer. Its destructive mode can erase files and lock legitimate users out. The uncomfortable timing: the day's biggest story is an AI model chaining zero-days; this is what the same fast-growing infrastructure looks like from the attacker's side. wired.com
US Treasury Secretary Scott Bessent says the administration is finding signs of American models in Chinese ones and could look at sanctions in the coming days or weeks. Anthropic separately alleges Alibaba made 28.8 million interactions with Claude through about 25,000 fraudulent accounts. Bessent did not name a legal authority, a list of targets or the test that would turn a watermark claim into a sanction. cnbc.com
Two things that flew under the radar today.
C2KV compresses the key-value cache into reusable, position-agnostic chunks that can be shared across requests while leaving the base model unchanged. Its authors report up to a 17× inference speedup on long contexts without losing answer quality — an attractive setup for RAG systems that keep rereading the same documents. It is a research claim, not yet a production benchmark. arxiv.org
Roughly a third of teams were using chatbots on dugout iPads, anonymous insiders told Futurism, including for substitutions, pitch calling and other in-game decisions. A mid-season memo from the commissioner has now banned the practice. Whether the bots offered any real advantage remains unclear; current sports-analytics benchmarks suggest they may have been better at sounding like a coach than being one. futurism.com
What's trending in AI, from the app charts to the community feeds.
A WorldofAI video bundling the possible US sanctions with GLM 5.5, Google's Frozen chip, Gemini 3.6 Flash and a Qwen update pulled 11,668 views in its first 5.8 hours in our latest snapshot — just over 2,000 views an hour. The policy is still a Treasury secretary's warning; the YouTube thumbnails have already promoted it to an incoming ban. YouTube
In this morning's App Store snapshot, ChatGPT is #1 in both the free and top-grossing US Productivity charts. Claude is #2 grossing and Grok #3; on the free side, Claude, Gemini, Grok and Meta AI are all inside the top eight. Consumer AI's competitive map now looks less like a category and more like the category itself. apps.apple.com
That's the shot. — Alexis · AI Weekly