Good afternoon. Today’s digest covers four stories spanning the full width of the AI world right now — from the Pentagon’s extraordinary allegations against Anthropic, to OpenAI’s audacious new research agenda, to the quietly unsettling economics of AI data collection, and a courtroom verdict that could cost one of Silicon Valley’s most powerful figures billions. Let’s get into it.
1. Anthropic vs. the Pentagon: Did a Leading AI Lab Become a “Supply-Chain Risk”?
In a remarkable legal and political clash that underscores how fraught the relationship between AI companies and the US military has become, the Department of Defense this week accused Anthropic of being capable of sabotaging its own AI tools during wartime operations. The allegation appeared in a court filing as part of an ongoing dispute between Anthropic and the Trump administration over the limits of how Claude — Anthropic’s flagship AI model — can be used in national security contexts.
Defense Secretary Pete Hegseth had already labeled Anthropic a “supply-chain risk” earlier this month, a designation typically reserved for foreign adversaries or compromised vendors. Now the Pentagon has gone further, asserting in legal documents that the company could theoretically manipulate deployed models mid-operation — potentially during live military missions.
Anthropic pushed back hard. Thiyagu Ramasamy, Anthropic’s head of public sector, filed a sworn declaration stating: “Anthropic has never had the ability to cause Claude to stop working, alter its functionality, shut off access, or otherwise influence or imperil military operations.” The company argues that once a military client is running a deployed version of Claude on their own infrastructure, Anthropic simply doesn’t have the technical access required to alter anything — before or during operations.
The dispute traces back to the fundamental tension at the heart of responsible AI deployment: model providers want to maintain safety controls and the ability to update or restrict their systems. The military, understandably, doesn’t want mission-critical tools that a private company in San Francisco can switch off or alter. These two positions may be genuinely irreconcilable — and the courts are now being asked to decide where the line sits.
Read the full Wired report on the Anthropic–DoD clash.
Why it matters: This case is a preview of a governance crisis that’s only going to intensify. As AI systems get embedded deeper into defense infrastructure, the question of who ultimately controls those systems — the vendor or the customer — has life-or-death implications. Anthropic’s position is principled on safety grounds; the military’s concern is operationally legitimate. There’s no clean answer, and that’s exactly the problem. The outcome of this case will shape how every AI company negotiates with government clients for years to come.
2. OpenAI’s New North Star: A Fully Automated AI Researcher by 2028
OpenAI has set itself a new grand challenge, and it’s arguably the most ambitious framing of the AGI question the company has ever committed to in public language. According to an exclusive conversation between MIT Technology Review and OpenAI chief scientist Jakub Pachocki, the company is now laser-focused on building a fully automated AI researcher — a multi-agent system capable of independently tackling scientific problems too large or complex for any human team to handle alone.
The roadmap is unusually specific. By September 2026, OpenAI plans to deploy an “autonomous AI research intern” — a system that can take ownership of a small number of well-defined research problems with minimal human guidance. That intern is the stepping stone to a far more capable system slated for 2028: a fully autonomous multi-agent research platform OpenAI claims will be able to generate new mathematical proofs, explore conjectures in physics, advance biology and chemistry, and even tackle complex policy and business dilemmas.
Pachocki described this agenda as OpenAI’s “North Star” — the unifying goal pulling together the company’s ongoing work on reasoning models, agentic systems, and interpretability research. The framing is significant because it’s not simply about making a smarter chatbot; it’s about building a system that does science — potentially accelerating discovery in ways that dwarf what any individual human researcher could achieve in a lifetime.
OpenAI faces stiff competition from Anthropic and Google DeepMind, and setting a concrete public milestone is as much a competitive signal as it is internal direction-setting. The September 2026 research intern target will be publicly tested — and publicly judged.
For developers who want to experiment with cutting-edge AI APIs as these models evolve, OpenRouter provides a unified gateway to the latest models from OpenAI, Anthropic, Google, and more — ideal for prototyping agentic workflows as the research-automation wave builds.
Read the full MIT Technology Review deep-dive on OpenAI’s research agenda.
Why it matters: A concrete September 2026 deadline for an autonomous AI research intern isn’t marketing fluff — it’s a commitment that will be publicly tested. If OpenAI delivers even a partial version of what Pachocki describes, it will reshape the conversation about AI’s role in scientific discovery and dramatically accelerate pressure on every competitor. If they miss, it will fuel legitimate questions about the gap between OpenAI’s ambitions and its execution. Either way, the gauntlet has been thrown — and the clock is running.
3. DoorDash’s Tasks App: Gig Workers Are Now Being Paid to Train the AI
While the high-minded debate about AI researchers and autonomous science unfolds in research labs and boardrooms, something much more immediate — and arguably more troubling — is quietly taking shape in the gig economy. DoorDash, the delivery giant, has launched a new product called Tasks, which recruits gig workers to perform data-labeling and behavior-capture work for AI training purposes.
Wired’s Reece Rogers spent time using the app firsthand and published a first-person account that reads like a dispatch from a near-future that has already arrived. The tasks include recording videos of mundane everyday activities — doing laundry, scrambling eggs, walking through a park — while the app captures these recordings to train AI models on how humans perform physical actions. Workers are paid per completed task, in classic gig-economy micro-payment fashion: low rates, high repetition, maximum flexibility for the platform.
The layers of implication are worth unpacking. Most immediately, this is the latest iteration of the “humans training the machines that will eventually replace them” arc — except it’s now being industrialized through one of the world’s largest gig platforms. DoorDash is effectively discovering that its massive, geographically distributed workforce of contractors can be monetized for any micro-task requiring human presence and judgment, not just food delivery.
There’s also a data ownership and compensation question buried deep in the terms. Workers generate rich, real-world behavioral data under agreements that almost certainly grant DoorDash broad rights to use, license, and profit from that data. A worker might earn a few dollars for a task; DoorDash captures training data potentially worth orders of magnitude more to AI model developers and robotics companies.
Why it matters: DoorDash Tasks is a window into a future that’s already arriving — one where gig workers don’t just deliver food but serve as the human substrate for AI training at scale. The economic logic is ruthless: the same platform flexibility that enabled on-demand delivery enables on-demand behavioral data collection. The workers helping train today’s AI models are, in a very direct sense, contributing to the systems that will eventually automate their own jobs. Whether that’s a feature or a design flaw depends entirely on your vantage point — but it’s happening either way, with little public debate.
4. Jury Finds Musk Liable for Misleading Twitter Investors — Damages Could Reach $2.6 Billion
A California jury delivered a consequential legal verdict on Friday, finding Elon Musk liable for misleading Twitter investors in the lead-up to his chaotic $44 billion acquisition of the platform. The verdict, while not a complete win for plaintiffs on every count, formally establishes liability — and the damages figure lawyers are floating is staggering: potentially $2.6 billion.
The case centers on public statements Musk made throughout 2022 as he negotiated, agreed to, threatened to abandon, and ultimately completed the Twitter deal. In a series of tweets and podcast appearances, Musk loudly and repeatedly raised alarm about the prevalence of bot accounts on Twitter, raising doubts about whether the deal would close and what the company was actually worth. This uncertainty depressed Twitter’s share price during the relevant window, causing investors who sold during that period to take losses they claim they wouldn’t have taken absent Musk’s statements.
Plaintiffs brought a class action claiming the bot narrative was part of a deliberate, coordinated scheme to manipulate the stock price downward. The jury rejected that sweeping theory — declining to find a broader scheme. But they did find that Musk’s specific public statements were misleading and caused real, quantifiable investor harm. Liability on that narrower finding was established.
Damages have not yet been determined. Because this is a class action covering a large pool of investors across the relevant period, even a moderate per-share damages calculation can compound into a colossal total. Plaintiffs’ lawyers have cited $2.6 billion as the potential ceiling, based on the scope of the class and the magnitude of the stock price movement attributed to Musk’s statements.
Why it matters: Beyond the headline number, this verdict sends a message that has implications across the entire tech industry. Musk is one of the most market-moving individuals on the planet — his statements about Tesla, Dogecoin, Twitter, and xAI have historically moved those assets significantly. A jury has now formally ruled, in a court of law, that this power comes with legal accountability. For founders across tech and AI who routinely make sweeping public claims about their companies’ capabilities, timelines, and competitive position, this is a precedent worth internalizing: words — even tweets — have legal consequences.
That’s the afternoon wrap. The through-line across today’s four stories is accountability — AI companies being asked to account for what their deployed models can do in wartime, researchers being held to ambitious self-imposed timelines, gig workers being asked to render their every physical action for pennies, and a tech billionaire being held to account for his words. The AI era is entering a more serious phase. We’ll be here watching it unfold.
Image: AI-generated
What to Read Next
- Claude Code Cheat Sheet 2026: Master Commands, Shortcuts, and Workflows for Developers
- Morning AI News Digest — Tuesday, March 24, 2026
- Evening AI News Recap — Monday, March 23, 2026
- Afternoon AI News Digest — Monday, March 23, 2026
- Browse all AI Stack Digest articles
Bookmark aistackdigest.com for daily AI tools, reviews, and workflow guides.
This article was produced with the assistance of AI tools and reviewed by the AIStackDigest editorial team.