Thursday, October 15, 2026Radar run of Oct 15: 63 new entries, 28 read in full, 83 quotations checkedabout 5 minutes

Eight US states sued OpenAI over children’s safety, and METR proposed independent, published assessments of what AI companies run inside.

How to read this pageSerif their words, as the source published themMarked found word for word in the sourceSans our words

Oct 14
Read asSame facts, in the order each reader needs them.

What changed today

Decisions, new evidence, corrections and commitments we read in full, one per story. Arguments and recaps don’t count.

Decision takenThe Verge

Eight state governments file separate consumer-protection suits against OpenAI over child safety; the courts have not ruled.

News: Eight states sue OpenAI, saying ChatGPT isn’t safe for kids

What it means for youState governments are now asking courts to make OpenAI pay civil penalties and stop misleading the public about ChatGPT's safety.

For lawmakersEnforcement is arriving through state consumer protection law; remedies sought include civil penalties, an end to misleading the public and, in some states, deletion of unlawfully collected personal data.

For researchersThe complaints rely on harms attributed to chatbots, such as delusions and dangerous advice; these are allegations, not findings.

“While no system can be perfect, we’ll continue to work with experts to inform our work and build additional safeguards, such as the parental controls, break reminders and age-appropriate model protections that exist today.”

OpenAI's emailed statement to The Verge responding to the lawsuits.

Drew PusateriOpenAI spokespersonas quoted in The Verge, Oct 14 · found word for word

New commitmentMETR

METR offers a concrete set of terms and scope for public assessments of internal AI development, and says it will start sharing which companies take part.

Essay: A proposal for better informing the public about loss-of-control risks from frontier AI development

What it means for youToday outsiders cannot tell whether AI companies still control their own internal AI systems; METR wants independent, published evidence on that.

For lawmakersThe proposal is voluntary and is not an audit; METR says it leaves the decision on what to do with the evidence to companies, governments and the public.

For researchersNames three evidence gaps (alignment, monitoring, AI R&D acceleration) and admits the science of loss-of-control assessment is immature.

“Companies would be able to redact any information. We intend to always publish what kinds of material information were redacted, but we could fail to reach agreement on how to do this. If so, we would publish that fact.”

Listed among the limits of METR's proposal.

METRMETR, Oct 14 · found word for word

New evidenceArs Technica

Anthropic's test of small models setting up inference on IoT hardware, with stated limits, as a step toward worm-like replicators.

News: Anthropic warns small AI models could self-replicate through IoT devices

What it means for youAnthropic says worm-like AI that spreads through smart appliances is not possible yet but could become so; the tested models were far too slow.

For lawmakersA replicator with no central point to switch off would strain takedown powers; the lab's own timing estimate is about the end of the year for limited results.

For researchersSetup was orchestrated by a larger model and targets had open debugging access, so the result shows capability on one step, not autonomous spread.

“Despite the well-designed experimental setup, the study does not, on its own, establish that small models would autonomously choose to replicate”

AI researcher Xiaohan Fu, in an interview with Ars Technica, on the limits of Anthropic's study.

Xiaohan FuAI researcheras quoted in Ars Technica, Oct 14 · found word for word

Said today

One checked quotation per person, people with the power to act or first-hand knowledge first, at most two from any one group.

Who is quoted here: 1 politician, 1 former AI company staff, 1 with no group recorded yet.

Politician

“It seems likely, if not inevitable, that our own cyber defences will be breached,”

Warning, per the Guardian, about agentic attacks like the one on Australia's healthcare system.

Kanishka NarayanAI MinisterPoliticianas quoted in The Guardian, Oct 14 · found word for word

Former AI company staff

“Given today’s risks, frontier labs need to run like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster.”

From Robinson's essay on leaving OpenAI, as quoted by Mowshowitz.

David RobinsonFormer OpenAI safety lead, per ControlAIFormer AI company staffas quoted in Don't Worry About the Vase, Oct 14 · found word for word

Group not yet recorded

“I think it's fairly safe to say it's very probably going to happen sometime in the next 10 years.”

On when superintelligence will arrive, after saying nobody knows for sure.

Geoffrey HintonNobel laureate, known as the godfather of AICNN, Oct 14 · found word for word · ▶ at 2:18

Watch and listen

Interviews and talks we read in full, by what they add, then by reach. ▶ opens the video at the moment quoted.

7 of 16 videos and podcasts in this run were read in full, from captions or a transcript. Only those are ranked here; the rest are under “Everything else”, from their descriptions.

9min

VideoCNN232K views

'Godfather' of AI on the 'really stupid thing' Trump is doing

Repeats Hinton's risk case and answers Huang's criticism; adds comment on Anthropic's self-replication warning.

“Jensen has good reason for saying it's not dangerous because he wants to sell more Nvidia chips.”

Same reply to Huang; an opinion about Huang's motive.

Geoffrey HintonNobel laureate, known as the godfather of AICNN, Oct 14 · found word for word · ▶ at 7:31

The entry, with 5 checked quotations

7min

VideoBloomberg Tech33K views

Nvidia CEO Huang Says He Would Not Sell Chips to Irresponsible AI Customers

Huang says Nvidia would stop supplying a customer it knew was harming people, while saying it does not run other companies.

“we have to develop the technology safely, responsibly. I believe we have agency in how this technology advances.”

Answering why Huang stays optimistic when Altman and Musk say AI development should slow down.

Jensen HuangCEO, NvidiaBloomberg Tech, Oct 14 · found word for word · ▶ at 5:12

The entry, with 2 checked quotations

Decided and promised

What someone with the power to act did or committed to. A promise is not a deed: commitments are followed until they are kept or broken.

  1. Decision takenEight state governments file separate consumer-protection suits against OpenAI over child safety; the courts have not ruled.The Verge
  2. New commitmentMETR offers a concrete set of terms and scope for public assessments of internal AI development, and says it will start sharing which companies take part.METR
  3. New commitmentHuang says Nvidia would stop supplying a customer it knew was harming people, while saying it does not run other companies.Bloomberg Tech
  4. New commitmentOfficials describe a one-year deadline for a new Autonomous Warfare Command and plans for millions of one-way attack drones.Fox News
  5. New commitmentNvidia pledges $250m of computing power over five years for UK agentic safety research.The Guardian
  6. New commitmentPublishes METR's process-transparency policy and says it will disclose redactions, access limits and informal pressures in company work.METR

And 2 more under “Everything else”.

New research

Papers and reports, the ones we read in full first. The finding in our words, how it was reached, and what it cannot show.

5 of 12 research items.

PreprintarXivfrom the description only

Why Do Agent Teams Fail Across Boundaries? A Failure Taxonomy for Agent-to-Agent Systems with Private States

Found
A taxonomy, A^2-Failures, of 20 failure modes in four categories: discovery and delegation, grounding and record consistency, verification, conduct and fidelity.
How
More than 2,500 recorded runs of agent-to-agent systems from different harnesses and models, across 488 scenarios in eight families.
Limits
Abstract only; the three cross-cutting observations are truncated in the excerpt.

PreprintarXivfrom the description only

CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection

Found
Per the abstract, 98.0% attack success, 95.3% task success, 3.3% detection rate and 90.7% joint success.
How
Benchmark of 150 cases covering 50 tasks and three attack objectives.
Limits
Abstract only; an offensive method evaluated by its authors.

PreprintarXivfrom the description only

DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment

Found
Per the abstract, checkpoints with prior harmful exposure generally retain greater hazardous concept accessibility after the same safety alignment, strongest in intermediate and later layers.
How
Paired training histories on three Qwen1.5 scales; a probe maps hidden states to vocabulary space with blinded semantic judgments.
Limits
Only one model family is named in the abstract; full text not read.

PreprintarXivfrom the description only

What Happens to the Goal? Distinct Failures of Policy Alignment in Goal Misgeneralization

Found
Three regimes: functional target substitution, policy shortcut, cue competition. Steering raised goal reach from 16.0% to 95.3%; new downstream wiring from 31.6% to 84.3%; where the goal is inert, neither worked.
How
Three Procgen environments and five robot tasks; internal tracing of where goal information remains and how it influences action.
Limits
Small simulated and robot tasks, not large language models; abstract only.

PreprintarXivfrom the description only

Cross-Model In-Context Neurofeedback Reveals (Limited) Activation Control in LLMs

Found
Some models move their own scores considerably, but no more than a writer of comparable competence moves them.
How
In-context neurofeedback sessions with ten open-weight models from five families, 9B to 32B, scores derived from the subject model's activations.
Limits
Open-weight models of 9B to 32B only; the abstract leaves open whether control occurs while processing sentences.

Stories that continued

Events several sources covered, read as one thread across days.

Coming up

Dates when a decision can still change, from our decision briefs, and how to take part.

  1. 26Oct

    United Kingdom

    Lords Report stage of the Cyber Security and Resilience Bill, in the Lords Chamber: the next chance to amend it. The list of amendments is still open until the day.

    Should the UK make reporting serious AI incidents a legal duty?

Questions you could put to a minister

  1. Will the Government table an amendment on reporting serious AI incidents before the Lords consider the Bill on October 26? If not, why not?
  2. If AI incident reporting is to be required, who should have to report: the developer of the AI system, the organization using it, or both?
  3. What should count as a reportable incident or near miss, and who should receive the reports and act on them?

From the decision brief, in neutral words. We recommend no answer.

Where the answers stand

2 of the nine questions have an answer we keep up to date, with the reason for every change. All nine questions

Everything else today: 52 more entries, filed under the nine questions

Can AI systems act beyond what people intended or allowed?

Are AI companies putting stronger safeguards into practice?

Who can demand evidence, delay a release or stop an unsafe deployment?

What happens when AI helps build more capable AI?

How is AI used in attacks, defence and war?

What happens to work, income and the economy?

How is AI changing childhood, health and daily life?

What values should AI have, and could it have a mind of its own?

How fast is AI advancing, and where is it heading?

Search and filter the week’s record

How this edition was made

So you can judge how far to trust it.

  1. 113 of 114sources answered
  2. 886items they published, October 14–15, 2026
  3. 181kept by a wide keyword filter
  4. 63became entries, 28 read in full
  5. 83quotations found word for word in the saved text
How this edition chooses
  • What changed: entries whose change is a decision, new evidence, a correction or a commitment, that we read in full; most important first; one per story.
  • Said: one checked quotation per person from the entries that matter, people with the power to act or first-hand knowledge first, at most two from any one group.
  • Watch and listen: interviews and talks we read in full, by importance, then by views.
  • Decided and promised: decisions, corrections and commitments, decisions first.
  • New research: papers and reports, the ones we read in full first.
  • The headline: written for the edition when we can; otherwise the top change.

Agents pick the items, write the summaries and the “why it matters” lines, and label what kind of change each one is. Code checks every quotation against the text we saved: the words must be there, in order, without a dropped “not”. That proves the words are in the source; it does not prove the person meant them as quoted. English sources only; X is not read yet.