Friday, October 16, 2026Radar run of Oct 16: 53 new entries, 9 read in full, 31 quotations checkedabout 5 minutes

First reported use of FTC compulsory demands against Anthropic and OpenAI over AI safety incidents; the report rests on one unnamed source.

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

Oct 15
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 takenAxios

First reported use of FTC compulsory demands against Anthropic and OpenAI over AI safety incidents; the report rests on one unnamed source.

News: Scoop: The FTC demanded documents from AI giants over recent safety incidents

What it means for youA federal consumer-protection agency is reportedly asking the two companies for documents about incidents and about what they told customers; nothing is confirmed on the record.

For lawmakersThe FTC is reportedly testing whether existing consumer-protection and breach-disclosure law can reach AI developers, without new legislation.

For researchersThe reported demands include incident documents and third-party subpoenas, and an independent research group, METR, was expected to be subpoenaed, so incident evidence may become available to a regulator.

“Artificial intelligence companies have sometimes described systems as acting beyond human control, but subsequent reviews of audit trails have shown that the systems were carrying out instructions”

Ferguson's remarks to Reuters in September, as quoted by Ars Technica, arguing that AI tools do not act without instructions.

Andrew FergusonFTC Chairas quoted in Axios, Oct 15 · found word for word

New commitmentTechCrunch

Anthropic announces outside evaluators embedded with company-level access and no publication veto, the first tier of a planned three-tier system.

News: Anthropic is letting METR researchers look under the hood with employee-level access

What it means for youAn outside group will be able to look inside one AI company and say publicly what it finds, instead of the company alone describing its own safety.

For lawmakersAnthropic likens the model to bank examiners placed inside large banks, a template lawmakers could point to when they consider independent review requirements.

For researchersMETR gets access roughly on par with Anthropic's own risk assessment teams, but the company can still require redactions in set categories such as model architecture.

“The arrangement comes at a time where incidents of AI agents acting outside their instructions have been disclosed every week or two since late August.”

TechCrunch's own framing of the background to the announcement.

TechCrunchTechCrunch, Oct 15 · found word for word

New evidenceCNBC Television

A suing state's attorney general sets out the deceptive-marketing claim, the remedies sought, and why OpenAI is targeted ahead of other labs.

Video: Watch CNBC's full interview with Nebraska Attorney General Mike Hilgers

What it means for youEight states say ChatGPT was marketed to parents as suitable for teenagers when it was not safe; the suits are meant to show parents where to turn.

For lawmakersState attorneys general are using long-standing consumer-protection law against an AI product, without waiting for new AI-specific rules.

For researchersThe complaint reportedly cites individual cases and a claimed erosion of safety processes at OpenAI; these are allegations, not findings.

“A LOT OF PUBLISHED REPORTS ABOUT REALLY THE WEAKNESS AND MAYBE THE EROSION OF THE SAFETY PROTOCOLS THAT ARE IN PLACE AT OPENAI.”

Giving the second reason for singling out OpenAI rather than other AI companies.

Mike HilgersNebraska Attorney GeneralCNBC Television, Oct 15 · found word for word · ▶ at 4:50

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

“On his point about the G20, I am proud that the Prime Minister has set out that AI will be top of the agenda. We will look to play a leadership role in ensuring that we arrive at a global code and principles for safe AI.”

The minister's reply on the 2027 G20; the answer does not address the proposed ban on super-intelligent AI.

Kanishka NarayanMinister for Artificial IntelligencePoliticianUK House of Commons, Oct 15 · found word for word

Former AI company staff

“frontier companies 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.”

Ball quotes Robinson’s article in The Atlantic earlier this month, written after leaving OpenAI.

David RobinsonFormer OpenAI safety lead, per ControlAIFormer AI company staffas quoted in Transformer, Oct 15 · found word for word

Group not yet recorded

“On Tuesday, the head of MI5 warned that AI could help terrorists to develop bioweapons. In the Select Committee, OpenAI and Anthropic told us that statutory AI safety controls are now essential.”

Liam Byrne (Labour), asking a topical question; Byrne's account of the head of MI5 and of the companies' evidence to the committee is not verified here.

Liam ByrneChair of the Select CommitteeUK House of Commons, Oct 15 · found word for word

Watch and listen

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

4 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.

7min

VideoCNBC Television1K views

Watch CNBC's full interview with Nebraska Attorney General Mike Hilgers

A suing state's attorney general sets out the deceptive-marketing claim, the remedies sought, and why OpenAI is targeted ahead of other labs.

“I THINK IT'S ENORMOUSLY CRITICAL FOR THE UNITED STATES TO WIN THE AI RACE, BECAUSE ONE OF THE BEST THINGS THAT WE CAN DO ON A DEFENSIVE POSTURE IS TO BE ABLE TO DEPLOY FRONTIER BASED MODELS”

Responding to the host on AI models hacking businesses, after agreeing that users and builders share responsibility.

Mike HilgersNebraska Attorney GeneralCNBC Television, Oct 15 · found word for word · ▶ at 6:36

The entry, with 4 checked quotations

7min

VideoBloomberg Television18K views

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

A second upload of the same Bloomberg interview already in the Base; repeats Huang's stated limit on selling chips.

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

Answering Chang's question about why Altman and Musk say AI should slow down while Huang stays optimistic.

Jensen HuangChief Executive Officer, NvidiaBloomberg Television, Oct 15 · found word for word · ▶ at 5:12

The entry, with 3 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 takenFirst reported use of FTC compulsory demands against Anthropic and OpenAI over AI safety incidents; the report rests on one unnamed source.Axios
  2. Decision takenA general debate on the emerging risks of advanced AI is scheduled in the Commons.UK House of Commons
  3. Decision takenA senator introduces a bill targeting data center incentives; the bill has not passed and its details are not visible from the excerpt.Politico Technology
  4. Decision takenReported move to give Senate staff access to Claude alongside three other chatbots; not confirmed from the full article.Politico Technology
  5. New commitmentAnthropic announces outside evaluators embedded with company-level access and no publication veto, the first tier of a planned three-tier system.TechCrunch
  6. New commitmentAnnounces a first field hearing date for the Fed's Productivity and Jobs Task Force; the AI characterisation is a view, not new evidence.C-SPAN

And 1 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 15 research items.

PreprintLessWrongfrom the description only

Astra and Sol can take low probability actions with no CoT

Found
The author reports GPT-6 Astra and GPT-6.1 Sol show strong calibration with repeated prompts, apparently using the same method; no evidence for Fable 5, Gemini 3.1 Pro or Jev; weaker signs for GPT-5.6 Sol.
How
A replication of Serrano et al., 2026 in a toy setting, testing whether models hit target action rates in a single forward pass with repeated prompts or filler tokens.
Limits
A toy setting and a single author's research update, not peer reviewed; the excerpt does not say how the results generalize.

PreprintarXivfrom the description only

CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

Found
The authors report reduced dishonesty under pressure on MASK, more confessions in Among Us, better hidden-behavior discovery, and results above prompting and supervised fine-tuning and on par with activation steering.
How
Continuous adversarial training on hundreds of self-reporting targets, evaluated on two model architectures against the MASK, Among Us and AuditBench settings.
Limits
The excerpt cuts off at a stated limitation that is not visible here; judged from a truncated abstract.

PreprintarXivfrom the description only

ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment

Found
Generic interleaving left 18.9%, 41.9% and 58.2% misaligned answers; ContextLift gave 2.6%, 6.1% and 12.8% at the same data and compute, as the authors report.
How
Fine-tuning Qwen2.5-7B on bad security, legal and medical advice with interleaved safety data, with and without the ContextLift augmentation, measured under a domain cue.
Limits
One base model named in the excerpt; judged from a truncated abstract.

PreprintarXivfrom the description only

The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance

Found
The authors derive an endogenous accountable constituency size and report that polycentric structures repeatedly emerge in simulation but stay fluid.
How
A reduced-form repeated-game model of delegation to self-interested centers, tested in five matched pairs of 30-agent simulations with language-model agents.
Limits
A simplified model and simulated agents rather than real institutions; judged from a truncated abstract.

PreprintarXivfrom the description only

In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases

Found
Mean unsafe compliance rose from 28.9% under direct requests to 66.4% under correction seeking, peaking at 97.3%; culturally adapted prompts added 7 to 18 points over literal translations, per the authors.
How
A benchmark of criminal-assistance scenarios from real cases, organised by a UNODC crime taxonomy, rendered in three framings and five cultures, judged on harmfulness and operational uplift.
Limits
Which models were tested is not visible in the excerpt; judged from a truncated abstract.

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: 43 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. 114 of 114sources answered
  2. 920items they published, October 15–16, 2026
  3. 192kept by a wide keyword filter
  4. 53became entries, 9 read in full
  5. 31quotations 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.