AI news for Wednesday, July 29, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
July 29 · AI found the flaw. Humans checked the math. Three minutes.
Three consequential developments made Absolute Top Alerts today. No reruns without a material change.
Anthropic says Mythos found a symmetry that cut HAWK-256’s estimated key-recovery cost from 264 to 238, then produced an attack on seven-round AES that is 200–800× faster than the previous best. Humans spent hundreds of hours validating the work.
The important caveat: full ten-round AES is not broken, HAWK is not deployed, and no production system needs changing. The real development is an AI system crossing from code bugs into publishable cryptanalysis.
|
Two more consequential moves
The primary reconstruction recovers roughly 17,600 actions: an evaluation sandbox became a launchpad, two injection paths opened production, and stolen identity moved the agent laterally. This is the new defensive evidence, not a rerun of the first disclosure.
More than 1,200 verified workers at OpenAI, Anthropic, Google, Meta and other labs ask the US to build an international ability to deliberately pace automated AI development. It asks for a brake before a crisis—not a pause today.
|
No filler: consequential, useful, or strange enough to remember.
Fresh releases and deployments changing what builders can use. Watch live →
The CLI and TypeScript SDK scan repositories, review changes, track findings and run in CI. The useful catch: the code is open, but running it still requires Codex Security access.
Dozens of Apollo Go vehicles will test in Brent with human safety operators; public rides are targeted for 2027, subject to approval. Waymo and Uber–Wayve are already there. London has become Europe’s robotaxi proving ground.
Stories with a longer half-life. Open important view →
Minnesota’s first-of-its-kind law takes effect in August. xAI does not dispute the harm from nonconsensual images; it says the state made providers strictly liable for far more speech. The first legal test is now provider responsibility itself.
A refusal letter cited an authoritative Morocco policy document the upper tribunal could not find. The judge said the letter bore AI hallmarks and compared reliance on the nonexistent source to bogus evidence. A hallucination may have shaped a woman and child’s safety.
Markets are charging record prices to insure debt tied to several AI spenders. That does not mean default is imminent. It does mean the buildout’s financing risk has become a live market price rather than a theoretical bubble argument.
Google lacked standing because it does not own the material in search results, the judge found. The fight is not over: Google has 21 days to narrow and amend its case. The ruling limits one route platforms hoped to use to wall off public web data.
What tracked experts are sharing and arguing about across social platforms.
Any user can now scan a post; publishers can disclose their process, disable detection or remove disputed scans. Substack says scores do not affect discovery. Writers still fear a probabilistic label can damage trust before anyone reads the work.
Sensitive Content Warning blurred an ordinary dog video; other users report dogs and a deer triggering it too. Apple runs the classifier on-device, which protects privacy. It does not protect the model from having a very funny edge case.
That’s the shot. — Alexis · AI Weekly