A judge told the Pentagon it can’t punish an AI company for saying no. A hundred of those companies spent the same week warning that the attacks are about to get worse. And the marketplace that taught machines to be smart quietly set a closing date.
1. A federal judge blocked the Pentagon’s blacklisting of Anthropic
This is the story of the week, and it isn’t close. U.S. District Judge Rita Lin ruled in Anthropic’s suit against the Pentagon, Defense Secretary Pete Hegseth, and several other agencies over the decision to label the company a national security supply-chain risk. She ordered the government to rescind every directive it issued against the company.
The 59-page opinion is blunt. Lin found the designation was unlawful retaliation against Anthropic’s First Amendment rights and that the government denied the company the due process the Fifth Amendment requires. “The empty invocation of national security is not a blank check to punish and retaliate against government critics,” she wrote.
How it got here matters more than the ruling. Anthropic was, at the time, the only major AI company working with the Pentagon on classified systems. Negotiations reportedly broke down when officials pushed for contract language letting the technology be used for “any lawful purpose.” The sticking points were domestic surveillance and autonomous weapons. When the talks collapsed, the administration reached for a designation normally aimed at foreign companies considered a threat to national security. The government can still appeal.
Why it matters: every AI company writes a usage policy. This is the first time we’ve seen what happens when one enforces it against the customer with the largest budget and the most leverage. The answer, for now, is that a court will back the company. If you’re picking a vendor for anything sensitive, the question “will this company hold a line?” just stopped being a branding question.
Coverage: CNBC, Axios, NBC News, Forbes.
2. Over 100 companies signed the same warning: the cyberattacks are coming
On August 27, more than a hundred technology and security companies put their names on one open letter about AI-enabled cyber threats. The signatory list is the part to read: OpenAI, Anthropic, Google, and Microsoft alongside CrowdStrike, Okta, Fortinet, AWS, Cloudflare, Cisco, Palo Alto Networks, IBM, Oracle, Check Point, and Hugging Face, which spent July on the wrong end of an agent incident we covered in the August 13 roundup.
The language is not hedged. “In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable,” the letter says. “The companies and public services our communities depend on — from hospitals to water treatment plants to the infrastructure that powers the internet — are at risk.” The ask is a “collective response” and new partnerships to raise security standards.
Each of the big labs now has a defensive program to sell you: OpenAI’s Daybreak, Anthropic’s Mythos, Microsoft’s Perception platform. Read the letter with that in mind. It is both a real warning and a market being created.
Why it matters: nothing in that letter is aimed at a two-person business, and that’s the point. The hospitals get a task force. You get the same phishing email, written better. The four things worth doing this weekend: turn on passkeys or an authenticator app everywhere that offers it, stop approving login prompts you didn’t trigger, verify any payment change by voice on a number you already had, and assume a convincing email in your vendor’s exact writing style is now cheap to produce.
Coverage: TechCrunch, CNBC, Axios.
3. Mechanical Turk closes September 30, after 21 years
AWS is shutting down Amazon Mechanical Turk permanently on September 30, 2026. It stopped taking new customers back in July. SageMaker Ground Truth is going with it. Amazon didn’t give a reason.
Jeff Bezos once called MTurk “artificial artificial intelligence” — humans doing, for pennies a task, the labeling and judgment that software couldn’t. That work built a meaningful share of the training data the modern industry sits on. The ending has a certain symmetry: Swiss researchers found in 2023 that up to 46% of Turkers were already using AI models to complete their assignments. The platform was outsourcing itself back.
Two groups need to move before the deadline. Businesses still routing data work through it — there are insurance and travel companies in that bucket — need a replacement, and the AI-native labeling shops that ate this market are the obvious landing spot: Scale AI, Mercor, Prolific. And the people for whom this was flexible remote income, in some cases full-time, have about a month.
Why it matters: this is the clearest picture yet of what “AI took the job” actually looks like. Not a dramatic replacement. A marketplace that quietly stops being worth running, with 30 days’ notice.
Coverage: CNBC, TechRepublic, Quartz.
4. A third of American adults now ask a chatbot about their health
New Pew Research numbers, out August 25, from 3,488 U.S. adults surveyed June 22–28. Thirty-four percent have used an AI chatbot for at least one of eight health reasons. The breakdown: 28% for quick health information, 25% to understand what’s causing symptoms, 22% for low-cost information they’d otherwise pay for.
Adoption is not evenly spread. Asian American adults are at 56%, adults under 30 at 44%. Roughly 47% of users call the answers extremely or very helpful — but only 29% say they’re very comfortable handing personal health data to these tools. People are using something they don’t fully trust, which is usually a sign the alternative is worse. The alternative here is a copay and a three-week wait.
The mental health finding is the one to sit with. Asked about chatbots for loneliness, 39% of Americans say they do more to hurt people than help; 19% say the reverse.
Why it matters: the useful line isn’t “don’t.” It’s what for. Chatbots are genuinely good at translating the language — decoding a lab result, explaining what a prescribed drug does, turning your vague symptom description into the specific questions to ask. They are not good at telling you whether you’re fine. Use it to prepare for the appointment, not to skip it, and don’t paste anything you’d mind seeing in a breach.
We took the same approach to kids and teens: ChatGPT for Teens and what parents can actually control.
Source: Pew Research Center, plus the companion study on chatbots and loneliness.
5. The money moved toward agents, twice
Nvidia is in talks to invest in Perplexity in a round valuing it north of $30 billion, per The Information. That’s a 50%-plus jump from the $20 billion mark a year ago. The number underneath it is the interesting one: annualized revenue past $750 million, up from under $250 million at the start of the year, with much of the growth credited to Perplexity Computer — the cloud agent that runs computer tasks for professionals. Nvidia has held a stake since 2023, and Perplexity committed in July to running its agent workloads on Nvidia’s Vera CPUs.
Alibaba, meanwhile, closed a HK$80 billion (about $10.2 billion) placement on August 26 — 710 million new shares at HK$112.70. It’s the largest primary follow-on ever by a Hong Kong-listed company and the world’s third largest this year, behind only Alphabet and Intel. All of it goes to AI: roughly 60% to global computing infrastructure, 40% to hyperscale data centers and upgrading cloud infrastructure toward what Alibaba calls an Agentic Cloud.
Why it matters: both bets are on agents doing work, not on chatbots answering questions. Perplexity’s revenue tripled on the back of an agent product. Alibaba is pouring ten billion dollars into the plumbing to run them. That’s the same shift we keep writing about from the small end — hire the teammate, don’t prompt the chatbot — showing up on the capital side.
Also this week
- Governance for agents is starting to take shape on three tracks at once: Google’s Agent Payments Protocol (AP2), NIST concept work on agent identity and permissions, and the AI AGENT Act (S.5051) in the Senate. Nothing is binding yet. It’s worth watching before you hand an agent a card number.
- More than a dozen state Attorneys General have asked OpenAI for answers on the Hugging Face breach.
The through-line: this was a week about who gets to say no. A judge said an AI company can refuse the Pentagon and not be punished for it. A hundred companies said out loud that the attack surface is about to get much worse. Amazon said a 21-year-old human labor marketplace isn’t worth running anymore. And the money went, in two very large chunks, toward agents that act instead of models that answer.
If you only do one thing: spend twenty minutes on the security list in section 2. It’s the only item here that touches your business this month.
If you only do one other thing: if you’ve been asking a chatbot about symptoms, keep doing it — but start pasting the answers into a note titled “questions for my doctor” instead of treating them as the end of the conversation.
Roundup for August 24–29, 2026. Sources linked inline.
Share this article
Leave a Reply