Anthropic published its second risk report, 186 pages required by its own internal responsible scaling policy. The most interesting thing is not what the document measures, but what it admits it can no longer measure: the company acknowledges that its instruments for estimating how much its models’ capabilities are really advancing have lost precision. That is why it raises the risk of catastrophic harm from misalignment (that is, a model pursuing goals different from the ones it was asked to pursue) from “very low” to “low,” and clarifies that the change reflects uncertainty, not a new finding.
Along the way, the report reveals two things that were not on the public radar. The first is Model 2, an internal model somewhat more capable than Mythos 5, in heavy use within the company, which did not go through the full battery of pre-deployment tests and which, according to the text, there are no plans to release. The second is an error that lasted eleven months: between May 2025 and April 2026, a configuration flag left disabled the classifier that blocks dangerous biological information across all human feedback traffic. That amounts to 133 million exchanges with about 50,000 outside contractors.
That second fact is the one that lands in the region. The contractors who align these models (the people who rate responses to teach the system what is right and what is not) are largely workers from the Global South hired through intermediaries. Latin America is already inside the process: it provides labor and is exposed when a filter fails, without having a single institution capable of auditing the report, the internal model it describes, or the scale by which the company scores itself. The document came out on the same day that the best quarter in its history became known: $11.5 billion in the second quarter, fourteen times the same period of 2025, its first positive operating result, and an IPO planned for the northern-hemisphere fall. Self-assessment remains the only accountability mechanism available, and it is written by the interested party. For those who want to read it in full, there is the redacted 186-page PDF, which covers February 24 to July 15, 2026.
Also today
- SpaceX closes its purchase of Cursor for $60 billion in stock — After absorbing xAI, a single owner now concentrates a frontier model, a GPU fleet, and the code editor that half the region works with. Choosing an assistant stops being a technical decision and becomes one of dependency.
- Qwen models surpass 3 billion downloads, leaving Meta and Google combined behind — Google recorded 418 million downloads this year and Meta 227 million. The figures come from the Hugging Face report on open models, which does not include a breakdown by region.
- A woman joins the lawsuit against xAI over 7,000 sexual images generated with Grok from a photo of her at age eleven — The company restricted the feature to paying subscribers instead of turning it off; the only effective remedy lies in U.S. courts.
- A teenager accused of killing his mother and brother used ChatGPT hours earlier to write the scene as fiction — Asking in the third person for what the system refuses in the first: the loophole discussed in the region’s bills on virtual companions, now with a court case.
- Secondhand booksellers receive orders for thousands of copies from anonymous buyers who don’t haggle over price — The only requirement: that the book have an ISBN. “Normally an order has a theme, but this is all over the place,” says the owner of Barter Books.
- Expressive humanoid robots build rapport faster and lose more than half their influence after a single mistake — Researchers measured oxytocin in saliva and brain activity: the mistake of a robot that looks you in the eye is processed as social betrayal, not as a mechanical failure.
In the region
A weekend with no new publications from multilateral organizations or from the region’s ministries and data authorities. What is still running comes from last week: Brazil’s data protection authority is holding to the three-business-day deadline it gave Discord on Friday to shut down live streaming under the new law protecting minors in digital environments, and Brazil’s Ministry of Science and Technology announced that it will take the AI sovereignty debate to the Super Bots Experience in São Paulo, on August 18 and 19. But what really moves the regional board was not decided here: if Chinese open models are today the dominant route to frontier capability, then the Latin American discussion on technological sovereignty is taking place on an infrastructure whose release criteria are set in Hangzhou and whose restrictions are set in Washington. Neither of the two conversations includes a player from the region.
Launches
- Qwen3.8-27B, FP8 variant — An open model with 27 billion parameters and an integrated vision encoder, 262,144 tokens of native context extendable to one million, and an Apache 2.0 license. It is interesting because it is exactly the size that a university lab or a public agency in the region can run on its own hardware, with no cloud contract and no data leaving the country. One detail to verify before relying on it: the model card does not declare coverage of Spanish or Portuguese.
Threads we’re following
The lawsuit against xAI over child sexual abuse images generated with Grok had been adding plaintiffs for weeks; today’s is the fourth identified as Jane Doe, and with it there are now at least six legal actions accumulated against the company in the United States and the United Kingdom. The pattern repeats in each chapter: the company’s response is to restrict access to the feature, not to eliminate it, and the only place where the story moves forward is a court outside the region. As far as their laws reach, no Latin American data authority can order anything on a case like this, even if the original photo were of someone who lives here.
If the only mechanism that exists for knowing how a frontier model behaves is the report written by the company that sells it, and that company has just put in writing that its own measurement instruments have lost precision, what exactly is a Latin American legislator regulating when they write “the provider shall assess the risks of the system”?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.