The Signal — February 24, 2026
Anthropic just published the most detailed public accounting yet of industrial-scale model theft by Chinese AI labs. The rest of the day's news barely registered by comparison.
Anthropic Catches DeepSeek, Moonshot, and MiniMax Stealing Claude's Brain
Anthropic identified three Chinese AI labs running coordinated distillation campaigns against Claude, using roughly 24,000 fraudulent accounts to generate over 16 million conversations. DeepSeek, Moonshot, and MiniMax each targeted Claude's strongest capabilities: agentic reasoning, tool use, and coding.
The details are specific and damning. DeepSeek ran over 150,000 exchanges and used Claude as a reward model for reinforcement learning. Moonshot operated through cloud proxy infrastructure to disguise its traffic. MiniMax focused on coding tasks. All three used synchronized accounts, shared payment methods, and coordinated timing.
Anthropic's framing is explicitly political: distillation undermines US export controls on AI chips by letting foreign labs close the capability gap without building from scratch. The blog post argues that what looks like rapid Chinese AI progress is partly built on extracted American capabilities. The timing, landing squarely in the middle of the chip export control debate, is clearly intentional.
This is the first time a major US AI lab has publicly named specific competitors and provided operational details of distillation attacks at this scale.
Sources: Anthropic Blog
OpenAI Says Its Own Benchmark Is Cooked
OpenAI published a detailed post explaining why it will no longer report scores on SWE-bench Verified, the coding benchmark it helped create in 2024. The reason: the benchmark no longer measures what it claims to measure.
Two problems killed it. First, at least 59.4% of the tasks OpenAI audited have flawed test cases that reject correct solutions. Second, every frontier model OpenAI tested could reproduce the exact human-written bug fixes used as ground truth, meaning they've all seen the answers during training. Improvements on the leaderboard now reflect training data contamination, not actual coding ability.
The honest part: OpenAI is recommending the industry switch to SWE-bench Pro, a benchmark where OpenAI doesn't currently hold the top score. That's an unusual move. Most labs only retire benchmarks where they're losing, not ones where they're winning. Simon Willison noted the charitable read: OpenAI is choosing measurement integrity over bragging rights.
The broader lesson is familiar. Benchmarks that use public data inevitably get contaminated. The shelf life of any coding evaluation is now measured in months, not years.
Sources: OpenAI Blog
The Humans Hidden Inside Humanoid Robots
MIT Technology Review published an accountability piece on the gap between humanoid robot demos and the human labor propping them up. The story is built on specific reporting: a worker in Shanghai spent a week wearing a VR headset and exoskeleton, opening and closing a microwave hundreds of times a day to train the robot beside him. Figure AI partnered with Brookfield to capture movement data across 100,000 residential units. Delivery workers wore motion-tracking sensors so their movements could feed robot training sets.
Then there's tele-operation. Neo, a $20,000 humanoid from startup 1X shipping this year, will have human operators in Palo Alto piloting it remotely when it gets stuck. The company's founder said he's "not committed to any prescribed level of autonomy." If your home robot is being driven by a remote worker watching through its cameras, the privacy implications are obvious. And the economics look less like automation and more like gig-work wage arbitrage with a robot suit on.
Jensen Huang declared 2026 the "era of physical AI." The reality is that physical AI currently runs on the same invisible human labor that powered every previous round of AI hype. The difference is that this time, the humans are wearing exoskeletons.
Sources: MIT Technology Review
On the Editor's Desk
Twenty-four events came through the pipeline today. We published three and killed most of the rest.
The biggest kill was a Wes Roth YouTube video about METR's AI agent capability research. The underlying data is real and significant: AI agent task horizons are doubling roughly every seven months. But our pipeline grabbed the hype commentary video instead of the actual METR paper or MIT Technology Review's earlier coverage. We're not publishing someone else's reaction video as a source.
Three separate stories about humanoid robots crossed the pipeline (Honor at MWC, an AI Summit in Delhi, and the MIT Tech Review piece we ran). The hype cycle for physical AI is loud right now. We picked the one story that actually interrogated the claims instead of repeating them.
We also killed three Towards Data Science tutorials, a Reuters homepage scrape that somehow got ingested as an event, a market forecast press release, and a generic AI regulatory compliance listicle. The web scraping filters need tightening.
An Anthropic persona selection model research paper came through and passed editorial review, but didn't make the top three. Same for an incident where an OpenClaw agent went rogue on a Meta researcher's inbox. Both are solid stories; today's lead was just too strong.