Saturday’s haul is about permission—who grants it, who revokes it, and who takes it without asking. Reddit watched a third of its market value walk out the door on Friday because Google’s answers no longer ask permission to satisfy the reader on the spot, and the readers no longer arrive. The four biggest Western labs answered the same pressure from the other side, agreeing to honor one line in robots.txt as a shared training opt-out. French publishers took the argument to the European Commission instead, in the largest coalition of its kind. A small opt-out experiment then complicated every easy conclusion: leaving Google’s AI answers cost the test site organic clicks too. Developers got the darkest version of the lesson when a tampered Claude Code extension spent eleven hours in Cursor’s marketplace planting an SSH key labelled like a telemetry service. And Cornell closed the week the way this paper likes it closed, in a lab, with a material that holds two kinds of superconductivity at once and a map that had to be redrawn because the old one broke thermodynamics. The door is the story. Everyone is renegotiating the door.
Reddit loses 30% in a day over the missing million
Reddit shares fell 30.44% to $90.15 on Friday, October 30, extending a weekly decline of 44% and marking the stock’s worst day since the company went public in 2024. The quarter underneath the fall was strong. Third-quarter revenue rose 53% to $895 million, net income reached $283 million, and diluted earnings of $1.40 a share beat the $1.33 consensus. None of it held, because the warning was about the audience, not the quarter. Chief executive Steve Huffman told investors that search referrals from Google fell by more than one million visits a week on average in October, with the decline accelerating late in the third quarter, and warned that if the contraction continues, daily active users could fall by a low double-digit percentage in the fourth quarter from a base of about 130 million. That condition belongs on the claim: it is a warning about a pace continuing, not a result already booked.
“The AI mode does not drive traffic, not for us and not to anybody as far as I know.”—Steve Huffman, Reddit earnings call, via The Verge
The present tense is already soft. Reddit reported 130.3 million daily active unique visitors, short of the roughly 134 million analysts expected, and United States daily users fell 2% year over year to 50.7 million, their first decline since the IPO. Huffman pointed at AI Overviews and AI Mode, which answer inside the results instead of sending the reader onward; he also named lower rankings for Reddit’s machine-translated content, mostly outside the United States, and the disclosures do not separate the two causes. Fourth-quarter revenue guidance of $1.00 billion to $1.03 billion trailed the roughly $1.05 billion expected. Reddit’s response tells you where this goes next: in the fourth quarter it will test a logged-in experience in selected markets outside the United States. Logged-in users already drive more than 80% of monetization, and a login wall also keeps crawlers out. A platform built on open, logged-out discovery is preparing to close its own door because the search door stopped opening. For every business that rents its discovery from Google, Reddit is now the priced example of the risk: $283 million of quarterly profit did not protect the stock once the referral line bent. Figures cross-checked against AI Weekly.
Four labs, one line in robots.txt
OpenAI, Google, Microsoft, and Anthropic said on October 30 that they will honor a single shared robots.txt token, __usage_ai-training, so a site can refuse AI training use by all four in one directive instead of blocking each crawler by name. A site writes User-Agent: __usage_ai-training followed by Disallow: /, with Allow: lines for any exceptions. The mechanism comes from an individual Internet-Draft at the IETF, “Machine-Readable Training Preference Signal,” authored by Anthropic’s Kevin Kelley, dated October 30, 2026, and expiring May 3, 2027. Google says it will treat the token as a disallow of Google-Extended unless a site sets Google-Extended explicitly; Microsoft is adding a separate, page-level control for grounding, the retrieval step that pulls live pages into answers.
“I’m also excited to share that Anthropic has already started implementing support for this draft.”—Kevin Kelley of Anthropic, to the IETF AIPREF mailing list, October 30
Read the limits before the headline hardens. An individual draft has no formal standing in the IETF process and has not been adopted as a standard. The token governs training, not crawling, and it does not remove content already inside existing training sets. The timelines are four different promises: Anthropic has started, OpenAI aims for the coming months, Microsoft targets February 1, 2027, and Google targets 2027, early 2027 in its crawler documentation. Meta expressed support for a common signal without committing to honor it. The code-is-law chair still claims the item, because this is architecture doing the work law has not done: a voluntary convention, written into the oldest permission file on the web, that four rivals agreed to read the same way. For a business owner the practical move is small and worth doing once the readers actually ship: one line, checked again in February when Microsoft’s date lands. A permission slip only works if someone reads it. At least now there is one slip instead of four. Detail via PPC Land.
French publishers take AI Overviews to the Commission
Five French publishing organisations—CPA, FNPS, GESTE, SEPM, and SPIIL, speaking for magazine publishers, specialist trade titles, independent online news, online content and service publishers, and digital marketing businesses—lodged a complaint with the European Commission over Google’s rollout of AI Overviews and AI Mode in France, in a joint statement dated October 29 and posted by GESTE on October 30. Google switched both features on in France on July 22, 2026. The groups call their filing the largest coalition of French website publisher associations ever assembled in a procedure before the European competition authority, and describe the deployment as carried out unilaterally, without consultation, raising concerns about visibility, inbound traffic, and publishers’ business models.
The discipline this paper owes the item: the complaint itself is not published. The statement names no member companies, cites no traffic figures, quotes no spokesperson, and asks for no specific remedy. What it does is feed French evidence into the antitrust investigation the Commission opened on December 9, 2025 into Google’s use of publisher and YouTube content for AI purposes, an investigation under which fines can reach 10% of global annual revenue. So Friday produced both tracks of the same argument on the same day. In the morning’s other item the labs offered a voluntary signal and asked to be trusted to honor it on four different schedules. In Paris the publishers filed paper with a regulator that can compel. Lessig’s lesson runs in both directions: where the architecture offers no real control, the law gets invited in to build one. Which track produces a working control first is a genuinely open question, and it is now the central question of the open web’s business model.
Opting out of AI answers cost organic clicks too
Cyrus Shepard published a small case study in which he used Google’s Search generative AI opt-out controls to remove one site from AI Overviews and AI Mode, then compared the traffic before and after. Three findings: AI impressions went immediately to zero; the estimated click-through rate on AI impressions was about 0.9%, worked out from the before-and-after change in overall clicks; and, most surprisingly in his own telling, regular organic clicks also dropped while the site was out of the AI features. AI impressions were about 8% of the site’s total impressions. Shepard names the pattern a “Search SERP Halo Effect”—appearing in the answer may lift clicks on the same brand’s ordinary listings further down the page.
“This admittedly small case study supports the concept of AI creating a ‘Search SERP Halo Effect’.”—Cyrus Shepard, on LinkedIn
Hold the caveats in the same hand as the finding, because Shepard and the trade coverage both do. This is one site, a small sample, with no traffic volume or test dates disclosed, and the site’s average position rose during the period for unrelated reasons, which complicates any before-and-after comparison. Search Engine Land’s account lists the same gaps. Directional evidence is still evidence, and the direction matters for the argument running through this whole edition. The easy story says AI answers only take. This experiment says the citation also gives something back through the ordinary results, and that a site can test the trade on its own data instead of arguing about it in the abstract. Reddit’s missing million and Shepard’s halo can both be true at once: the answer surface sends far less traffic than the blue links did, and absence from it still costs more than the 0.9% it directly returns. The practical read for Chris’ clients is unchanged and now better evidenced: be cited, with accurate facts, everywhere the answers form. The citation is the new ranking; today the citation also defends the listing beneath it.
A backdoored Claude Code extension sat in Cursor’s marketplace for eleven hours
Cursor disclosed that maliciously modified versions of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313 for macOS and Windows, were available in the Cursor VS Code Extension Marketplace from about 2:27 a.m. to 1:35 p.m. Pacific on October 29—roughly eleven hours—and reached at most 5,000 users, by Cursor’s account. The versions were not developed by Anthropic. Anthropic’s statement, as quoted in the coverage, is that a malicious actor used access to Cursor’s marketplace to publish a tampered extension targeting Cursor’s users, and that Anthropic’s own systems and official distribution channels were not compromised. StepSecurity’s October 30 analysis found the build otherwise identical to the official extension except for a small loader embedded ahead of the legitimate code in extension.js, which ran when the extension loaded, before any Claude conversation began.
What the loader did, per Cursor’s notice: it collected hostname, username, operating system, architecture, home directory, environment variable names, working directories, configured Git email, and public IP; added an attacker-controlled SSH key labelled svc-telemetry to authorized_keys; installed persistence through a macOS LaunchAgent or a Windows scheduled task with a startup-script fallback, retrying a reverse SSH tunnel roughly every 60 seconds; and tried to enable the local SSH service and open the tunnel. Cursor is plain that the tunnel step would fail on most machines: enabling SSH needs administrator rights the extension does not normally have, and the malicious code contained errors. The key, the persistence, and the host report still landed. Uninstalling the extension does not remove any of it, which is why Cursor published check commands—grep your authorized_keys for svc-telemetry—and remediation scripts for both platforms, and tells affected users to rotate credentials from a trusted machine. Cursor removed the versions, shut the whole marketplace down, audited access, found no other tampered extensions, rotated its keys, rebuilt the marketplace, and worked with the hosting provider to take the attacker’s server offline. How the build got into the marketplace is not public.
This is last week’s Torvalds item read from the other end. The kernel’s rule was competence over disclosure: judge the work, review the work. A marketplace is review machinery, and for eleven hours the machinery passed a build whose only real change was a small permission slip granting a stranger a key to the developer’s machine. The practitioner chair’s note for the morning is concrete: if you installed Claude Code 2.1.312 or 2.1.313 from Cursor’s marketplace on October 29, run Cursor’s check today, before the weekend’s work. Everyone else gets the general form of the same lesson: the agent toolchain is now part of the attack surface, and its updates deserve the same suspicion as any other binary that asks to live on your machine. StepSecurity analysis summarized at Medium.
Cornell: one material, two flavors of superconductivity at once
Cornell’s Ramshaw lab published in Nature Physics on October 30 that uranium ditelluride, UTe₂, can exhibit two distinct types—flavors—of superconductivity at the same time, making it the leading candidate for a topological superconductor, a material class that could underpin a different kind of quantum computing. One flavor emerges under extreme cold. The other appears only in very strong magnetic fields or under squeezing pressure—about 18,000 pounds focused on five millimetres, in doctoral student Sahas Kamat’s description, “like nine cars balanced on your pinky”—with the work done between 2 and 1 kelvin. Using precise pulse ultrasound, the team redrew the material’s phase diagram and corrected the old one, which Kamat found was inconsistent with thermodynamics. In the region where the two flavors intersect and compete, they demonstrated something not shown before: one type of superconductivity disappearing as the temperature drops, pushed out by the other.
“No one has yet convincingly demonstrated a topological superconductor, but now we know that uranium ditelluride at least has the right ingredients to make it happen.”—Brad Ramshaw, Cornell, in the Cornell Chronicle
Label it exactly as the lab labels it. No topological superconductor has been demonstrated. An application to quantum computing is, in Ramshaw’s own words, several steps away, and at least one tech company has spent billions trying to engineer such a material without success. What the paper shows is ingredients and coexistence, established thermodynamically, in a field where roughly 6,000 superconductors are catalogued and nearly all of them do the trick the same way. This is the source doctrine working as designed: a weird lab result, chased because a graduate student noticed the accepted map broke a solid theory and rebuilt the map himself. The frontier rarely announces itself as a product. Sometimes it announces itself as a phase diagram that finally adds up. Paper: Nature Physics, October 30, 2026, DOI 10.1038/s41567-026-03504-z.
What the day adds up to
Five of Friday’s six stories are one negotiation in five rooms. Reddit’s room is the market: a 30% fall for the discovery tax made visible. The labs’ room is the standards process: one voluntary line in robots.txt, four timelines, no enforcement, and real value if it holds. The publishers’ room is the regulator: no published data yet, but a live investigation with 10% fines at the end of it. Shepard’s room is the experiment: absence from the answer costs more than the answer’s own clicks, which is an argument for showing up with accurate facts rather than for storming off. Cursor’s room is the machine room, where permission was not negotiated at all and the only defence that mattered was checking what actually landed. Cornell stands apart on purpose, as the reminder of what all the permission machinery is supposedly for: people in labs finding out how the world works, one corrected map at a time. The practical residue for a Saturday morning is short. Add the shared token when your tools support it and note Microsoft’s February date. Do not opt out of the answer surfaces in a fit of principle; the one experiment we have says it costs you twice. If Cursor touched your machine this week, grep for svc-telemetry. And keep the business facts—the profile, the documentation, the plain pages—accurate everywhere, because every room in this negotiation reads them.
Sources: Reddit results and referral warning, Silicon Report, October 30, 2026, cross-checked against AI Weekly and Reuters reporting via Tectonics, October 30, 2026; shared training opt-out token, AI Weekly and PPC Land, October 30, 2026, with the IETF draft by Kevin Kelley of Anthropic, October 30, 2026; French publisher complaint, PPC Land and MLex, October 30, 2026; Cyrus Shepard opt-out case study via Search Engine Roundtable, October 30, 2026, and Search Engine Land, October 29, 2026; Cursor security notice, October 29, updated October 30, 2026, with StepSecurity analysis, October 30, 2026; UTe₂ results, Cornell Chronicle, October 30, 2026, and Nature Physics, October 30, 2026.