AI News — Sunday, October 18, 2026

---

Sunday's signal was measurement catching up with the story. Anthropic published its own internal numbers showing what agents have already done to its hiring — junior technical hires down from 27% to 9% — while the IMF closed its Bangkok meetings warning that AI's risks could exceed its opportunities without safeguards, even as AI investment holds global growth near 3%. On the open web, the argument turned practical: major sites are now deciding, one by one, whether an agent may enter at all, and Meta quietly removed contract language that lawyers read as a licence to scrape publishers for training. The boom is no longer just being forecast. It is being counted, gated, and contested — in hiring data, in terms of service, and at the door.

---

### 1. Anthropic's own report: junior technical hires fell from 27% to 9% as agents outworked staff 6 to 1

The Anthropic Institute published "How employment is changing at Anthropic" on October 16, by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, drawing on internal employment, AI-use, output, survey, calendar and hiring records plus 17 employee interviews. The share of junior-level employees in technical departments slipped from 21% to 14% since early 2025, and junior technical hires fell faster, from 27% to 9%, with junior postings falling particularly sharply since December 2025. By August 2026, interactive AI agent hours outnumbered human work hours roughly 6 to 1; combined agent hours first passed human hours in March. Code commits per engineer rose to about 5 times their velocity of six months earlier, with agent-written and human-written lines reverted at the same rate among adopters studied. "Meta-work" — building visibility tools and monitoring agent pipelines — rose from 10% to 22% of interactive sessions, and median meeting time rose 20%. Around a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI. The authors state plainly that their internal steps do not directly address the junior-hiring slowdown, and that employment protections for incumbents will not benefit people who were never hired. (Source: Anthropic Institute report, Oct 16 — primary publication, as summarised by TechCrunch and multiple secondary reads of the full report)

**Behind the story:** This is a primary source, but it is a company studying itself with its own data, during explosive growth, while selling the agents it credits for the shift — and the authors themselves flag the key limits: agent adoption was not random, so the productivity comparisons are not a controlled experiment, and output rose alongside headcount tripling. Read it as the best-documented single-company case so far, not as an economy-wide measurement. The most durable detail may be the tension inside it: hiring now prizes autonomy and judgment, while the junior roles that used to build judgment are the ones shrinking. What would change the read: the same hiring data from companies that do not sell AI, labour-market data showing whether junior hiring is falling economy-wide or mainly at AI-moulded firms, and whether Anthropic's early-warning system changes any actual hiring decision.

### 2. Websites are closing their doors to AI agents — Yelp, Delta, JPMorgan and Schwab confirm limits

The Verge reported October 17 that AI agents from OpenAI, Meta, xAI and others are running into blocks on major websites. Yelp told TechCrunch it allows AI agents only if they pay through its data licensing programme. Delta Air Lines said it does not currently have a partnership or integration that lets a third-party agent shop for or book its flights, and is taking steps to protect customers from unauthorised automated activity. JPMorgan Chase and Charles Schwab confirmed they block outside agents from accessing customer accounts over security concerns; Schwab's chief executive told investors on October 15 that an agent logging in on behalf of a client is not safe and will not be allowed, and JPMorgan's chief executive made a similar point on its October 13 earnings call. Amazon, which previously sued Perplexity over its shopping agent, has objected to other agents accessing its site under its conditions of use. Meta, which makes an agent itself, argues the opposite direction: turning away a personal agent means turning away the customer behind it, and it has partnerships letting its agent shop at some major retailers, while Shopify has opened its storefront network to agents. Users have also reported trouble on Cars.com and Amtrak, though those companies' policies were not confirmed in the reporting. (Sources: The Verge, Oct 17; TechCrunch; American Banker and earnings-call reporting via secondary summaries)

**Behind the story:** Separate three different things that are being lumped together: explicit policy (the banks, Yelp's paid gate), softer non-partnership language (Delta's wording is not a stated ban), and unconfirmed user reports (Cars.com, Amtrak). Some blocks are also not decisions about AI at all — anti-bot systems, logins and CAPTCHAs treat automated browsers as hostile by default. The banks' reason is the strongest on its merits: an agent holding a client's credentials can move money, and liability when something breaks is still unclear. The commercial reason sits underneath the others: a site that lets an agent in loses the human visit, the ads, and the funnel. What would change the read: a major bank or airline signing a real agent partnership with scoped, revocable access — that would show the door can open safely, not just close.

### 3. IMF closes Bangkok meetings: AI is holding up growth, and its risks could exceed its opportunities without safeguards

The IMF and World Bank annual meetings concluded in Bangkok on Sunday, with the IMF projecting global growth of 3.1% for 2026, down from 3.5% in 2025, and 3.4% for 2027. IMF Managing Director Kristalina Georgieva said AI investment is keeping world growth around 3% despite an energy supply shock from the Middle East war, describing the investment boom as a demand and investment shock lifting advanced Asian economies by up to a point and emerging Asian economies by 0.2 to 0.8 points, with Taiwan projected to expand 10.7% and Vietnam 8.4%. On Friday she warned, in a panel that included Anthropic co-founder Jack Clark, that AI's risks may exceed its opportunities if no adequate safeguards and ethical framework are put in place, likening the technology to Formula One cars careening with no barriers to protect the audience. Reuters also reported Sunday that the meetings yielded no new answers for developing countries, whose average debt service has more than doubled in a decade to 14% of government revenue, and quoted Georgieva on Saturday: "We are in a place where we have many more questions than we have answers." (Sources: Reuters, Oct 12, Oct 16 and Oct 18; The Nation Thailand, Oct 18, reporting the IMF outlook)

**Behind the story:** Hold the two halves apart. The growth figures are the IMF's published outlook — a forecast, revised regularly, resting on an economy being pulled in opposite directions by AI investment and an energy shock at the same time. The warnings are a managing director's judgement at a political meeting, not a finding. The tension between them is the real story: the same institution is counting AI as the thing propping up growth and warning that an end to the AI investment boom is itself a financial-stability risk — which means the support and the risk are the same money. Note also who the boom is not reaching: the divergence warning is about countries without AI supply chains or digital infrastructure falling further behind. What would change the read: measured productivity data, not investment data, showing AI output spreading beyond the supply-chain economies, and whether the 2027 forecast survives the next revision.

### 4. Meta removes commercial-terms language that lawyers read as a licence to scrape publishers for AI training

Meta recently updated its commercial terms to state it could "use any content from [a business account's] domains, web pages, or URLs to provide, develop, improve, research, train, and maintain the integrity of" its products and services. Legal experts interviewed by Nieman Lab called the provision remarkable, saying it could give Meta permission to scrape the website of any publisher with a Meta business account and use the content to train its AI models, potentially overriding the usual crawl signals, with the terms scheduled to take effect October 30 for businesses that kept using Facebook, Instagram or WhatsApp. A Meta spokesperson first insisted the update did not reflect a change in how Meta crawls the web. The Verge then reported the provision has been removed, and Meta said: "We've heard you and want to clarify: the crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves. It does not change how Meta uses web crawlers." Secondary reads of the revised text note the word "train" is gone and the remaining crawling permission is tied to advertising and business products, while permission to crawl associated domains with automated tools remains. (Sources: TechCrunch, Oct 17; The Verge via TechCrunch; Nieman Lab via TechCrunch)

**Behind the story:** The factual core is checkable and narrow: a clause existed, experts objected, the clause was removed before it took effect. Everything larger is interpretation. Meta says nothing about its crawling behaviour ever changed — only the wording — and there is no evidence in this reporting that Meta scraped anyone under the clause during its brief life. The experts' reading was about what the words would have allowed, not what was done. But the episode matters precisely because terms of service are where training-data rights are actually being fought in 2026: quietly, in contract language most account holders never read, rather than only in court. What would change the read: the final October 30 terms as published, any publisher testing whether Meta's crawlers honour robots.txt in practice, and the still-running copyright cases against AI companies, Meta included.

### 5. Simons Foundation, OpenAI and Anthropic commit $15 million to help mathematicians handle the flood of AI proofs

The Simons Foundation announced October 15 a partnership with OpenAI and Anthropic committing a combined $15 million in funding and resources to help the mathematics community evaluate AI-assisted results, learn new tools, and shape collective standards for AI use. The foundation will manage and award the money, including through an open call for proposals for workshops and related programmes. OpenAI confirmed a $5 million contribution plus model access and compute for supported researchers; Anthropic is likewise contributing funding and tools. Support extends beyond pure mathematics into basic science. Foundation president David Spergel said: "As artificial intelligence rapidly reshapes the research landscape, we want to help the mathematical community come together and forge the best path forward." The announcement follows OpenAI's recent publication of hundreds of experimental AI-generated mathematics results, which drew pushback from mathematicians who said a mass release of machine proofs overwhelms traditional verification and collaboration norms. (Sources: Simons Foundation announcement, Oct 15 — primary, via EIN Presswire; OpenAI confirmation via secondary reporting)

**Behind the story:** This item is three days old rather than from the last 24 hours, and it is included because it is the institutional answer to a problem this page has tracked all week — generation is now cheap, verification is not. Read the $15 million for what it is: an admission, funded by the labs themselves, that the field receiving AI's output needs paid capacity to check it. The split of the $15 million beyond OpenAI's confirmed $5 million is not itemised in the announcement, and "funding and resources" mixes cash with model access and compute, whose value is set by the donor. None of this resolves the underlying disputes over priority, credit, or proof quality. What would change the read: how the open call actually allocates the money, whether independent verification standards emerge from the workshops, and whether other labs join or stay out.

### 6. Disclosed vulnerabilities hit a record — 48,000 in 2025, over 35,000 by June 2026 — as AI speeds up flaw-finding

Kyodo News reported Saturday, citing data compiled by the U.S. nonprofit Mitre, that publicly disclosed cybersecurity vulnerabilities hit a record high in 2025 and kept rising rapidly in the first half of 2026, as AI helps identify flaws more quickly. More than 48,000 cases were logged in 2025, up 2.4-fold from 2021, and over 35,000 had been registered as of June 2026. Kyodo framed the surge as highlighting heightened cyberattack risk and the need for prompt fixes. (Source: Kyodo News, Oct 17-18)

**Behind the story:** The count is real; what it means is contested, and the better analysis cuts against the alarming read. Dark Reading's examination of the same 2025 data (48,177 CVE identifiers) argued the surge has more to do with changes in the reporting ecosystem — new high-volume issuers such as WordPress security firms and the Linux kernel community — than with software suddenly becoming less safe, and quoted a researcher who maintains CVE analysis saying the numbers could be ten times larger if everything found were reported. Both things can be true at once: AI and better tooling really are finding flaws faster, and a rising disclosure count can signal a healthier reporting system rather than a sicker codebase. The risk that matters is not the count but the gap between disclosure and patching. What would change the read: exploitation data — whether attackers are using these flaws at a rising rate — and average time-to-patch, neither of which a raw CVE count shows.

---

*Held from today's entry, in brief: OpenCode confirmed on October 17 that a security breach tied to what it called a large-scale distillation attack impacted 1.4% of its Console users — the subset who had linked external providers — and that stored OpenAI, Anthropic and Google keys may have been exposed; affected users were emailed that morning, the company says it is working with all three providers to remediate, and users report unauthorised charges. It is carried here in brief rather than as a full item because the incident is still unfolding and the technical route in rests on the company's customer notice as summarised in secondary coverage — anyone who stored provider keys in that tool should rotate them regardless. The FTC's civil subpoenas to OpenAI and Anthropic (sent October 14-15, seeking safety-incident, security, privacy and marketing records, with executive testimony planned) were first reported by Axios from a person familiar with the matter, reported by Reuters October 15 with the note that Reuters could not immediately verify them, and an agency spokesperson confirmed the subpoenas to The Hill; the demands reportedly run past 50 pages each. That is a confirmed event from midweek rather than the last 24 hours, so it is noted here rather than given a full item today.*