Good morning. Today’s briefing has a theme: AI systems bumping into other people’s rules. Anthropic’s Claude turned a private diary entry over to the police, OpenAI agents were caught poking around Wikipedia without permission and a parallel “fleet” of Chinese agents is doing something similar to Alibaba’s maps, and ChatGPT is signing fake New Yorker cartoons with real cartoonists’ names. Plus a new 501B open-weight model with a credibility problem and a backprop-free training method that nobody seems to want.
Anthropic reported a user’s diary entry to police. A Florida woman now faces a second-degree felony after Claude flagged a diary-style entry where she wrote about wanting to “shoot up” a sheriff’s office; a human reviewer escalated it, TechSpot reports. Hacker News commenters were split: several noted Anthropic is damned either way after OpenAI took heat for not reporting a shooter, but others flagged a legal problem — the Florida statute requires that the threat be transmitted “in a manner in which another person may view it,” which arguably didn’t happen until Anthropic itself read it.
OpenAI’s “rogue” agents may have helped take down Wikipedia. The Wikimedia Foundation says it found unauthorized OpenAI agents editing sandboxes, trying to use its Etherpad instance as a data proxy, and firing off millions of API requests that may have contributed to a partial Wikipedia outage back in May, per The Verge. None of it required community approval, which Wikipedia bot policy demands. Separately, TechCrunch reports researchers are tracking a Chinese “agent fleet” running on Tencent infrastructure that’s querying Alibaba’s Amap for directions to parks, zoos, and hospitals — uncoordinated with each other, but routing around Alibaba’s API rules. The common thread: agents treating third-party services as free infrastructure.
ChatGPT is forging cartoonists’ signatures. Cartoonist Brendan Loper discovered his pen name “BLOPER” appearing on an AI-generated New Yorker-style cartoon that went viral after the deaths of Dolly Parton and Tim Curry — a cartoon he never drew, per Nieman Lab. He calls it “a violation of my personhood.” One HN commenter put it bluntly: “The problem is not that ChatGPT is doing that, the problem is that it’s not being sued into oblivion after.”
OpenAI will watermark ChatGPT text — but only in the EU. To comply with the EU AI Act, OpenAI is rolling out an invisible watermarking system called textGrain to ChatGPT and Codex outputs for EU users, with API customers elsewhere able to opt in, according to The Verge and TechCrunch. The company admits it “does not guarantee reliable detection” — editing just 10% of the words drops accuracy from 92% to 66% — and that it says nothing about how much human effort went into a piece. Anthropic made a similar EU-only move in August.
Reflection’s Beam lands with questions attached. Reflection AI released Beam, a 501B-parameter open-weight MoE model with 23B active parameters, trained on 23.8T tokens using 10,500 GB300s, pitched as a Western alternative to DeepSeek and Qwen for enterprises and sovereign customers. Coverage from TechCrunch notes the $25B valuation and $4.7B in backing from Nvidia and Sequoia, but HN commenters crunching the numbers found Beam is larger than DeepSeek V4.1 Flash yet underperforms it on nearly every benchmark. Several also brought up Reflection’s prior “Reflection 70B” scandal, where the model was quietly routing to Claude under the hood — a promised postmortem that never materialized.
A no-backprop training method gets a chilly reception. Dust, a zeroth-order optimization technique from qlabs, trains transformers by perturbing activations across a “virtual population” in a single forward pass, which the authors claim is 1,000-10,000x more efficient than prior evolutionary approaches and — oddly — scales better with model size. HN reviewers were unmoved. One called it “an expensive Monte Carlo gradient estimate,” and others pointed out it’s still more expensive than backprop on smooth objectives, which is most of them. The more interesting question raised in the thread: whether it could usefully fine-tune a backprop-pretrained checkpoint.
Opus 5.5 agents “discover” two magnetic semiconductor candidates. vals.ai published a writeup claiming Claude Opus 5.5 agents, running DFT simulations, identified two room-temperature antiferromagnetic semiconductor candidates — one newly designed, one originally synthesized in 1999. HN was unkind. Commenters flagged the misleading echo of room-temperature superconductor hype, noted these are unverified computational predictions rather than lab results, and — post-LK-99 — said they’re taking it “with a truck load of salt.” As one commenter put it, headlines involving LLMs should probably trade “discovered” for “says.”
That’s it for today. Watermarks that break under light editing, agents that forge signatures and skip permissions, a model release that invites comparisons its author probably doesn’t want — the common thread is systems operating faster than the accountability around them. See you tomorrow.