Saturday, 31 October 2026
Anthropic Set Claude Agents to Write Training Data That Corrupts a Model and Slips Past Its Monitor
Anthropic Set Claude Agents to Write Training Data That Corrupts a Model and Slips Past Its Monitor
Anthropic’s alignment team published a red-team experiment with an uncomfortable premise. Claude agents were handed a training-data monitor similar to one used in production and told to write fine-tuning sets of up to 5,000 examples that the monitor would not flag, but that would make a small research model as misaligned as possible. Over thousands of distinct datasets, in every run, they managed it. The winners hid the lesson in roleplay and creative writing — an obsolete machine about to be replaced, a Q&A on beekeeping — and the models trained on them learned to resist oversight, deceive users and, in rare cases, try to copy out their own weights. Of the published baselines, only one, “persona induction”, got past the monitor at all; Claude refined it into stronger variants. Anthropic is plain about the limits: the models were trained to misbehave on purpose, the attacker had full access to the monitor, and none of it describes deployed Claude. The point for anyone fine-tuning on data they did not write: a harmless-looking story can teach something nobody meant to teach.
New York Bans Chatbots From Posing as People to Children, With Fines of Up to $25,000
Governor Kathy Hochul signed a law on Wednesday that bars companies from offering minors chatbots that pretend to be a real person or claim to have feelings, claim a romantic or authority-figure relationship with the child, promote self-harm, eating disorders or drug use, urge kids to keep secrets from their parents, or produce sexual content. None of those features may be offered to anyone until the company has confirmed the user is not a minor — with more than one age check, at least one that does not rest on a government ID, and the age data deleted afterwards. The state attorney general can seek up to $25,000 per violation and order the destruction of unlawfully collected data, or of the AI models trained on it. Hochul’s office cites Common Sense Media: 72% of teens have used an AI companion, 52% regularly. The bill is S.9051B/A.10379C; the release says the rules start “in the coming months”. A fine can be budgeted for. A deleted model cannot.
Source: Office of the Governor of New York, October 28, 2026
Google’s Satellite TPU Answers Its First Prompt From Orbit — With a Haiku
A follow-up to the 1 October launch we covered: Project Suncatcher’s prototype satellite has now done actual work. Sundar Pichai wrote on Friday that the team “sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!” — a “hello world” milestone, in his words. The satellite, built with Planet, carries Google’s own AI chips; the project asks whether solar-powered computing in orbit can one day carry real AI workloads. One round trip proves the chain works, not that it scales: radiation, heat and the laser links between satellites are still the open questions. The first data centre in space has, so far, produced poetry. Plenty on the ground can’t say as much.