|
Also inside today's 13 stories
● Tampered Claude Code Extension Carried a Hidden Backdoor
● Library, Disability, and Tech Groups Back AI in NYT-OpenAI Suit
● After Ottawa's Stake, Cohere Pledges C$4 Billion and 1,500 Jobs
AI developments from the last 24 hours
For years, investors bet on how big AI would get. This week they started asking who pays for it. The Nasdaq fell 2.86% on Friday, its worst day since June, as traders dumped AI-related stocks, CNBC reports. It's heading for its worst month since March 2025. The Wall Street Journal called it the week the cracks in the AI trade grew deeper.
Earnings set it off. Amazon missed expectations, stuck with a $220 billion spending plan for the year and burned through $42.4 billion more cash than its business brought in. That came a day after Alphabet's worst day on record, when Google said its AI search earns less than the old kind (D.A.D., October 30). Intel plunged 29%, its worst day ever, on a weak outlook. Chipmakers followed: AMD fell 11%, Broadcom 5% and Nvidia 4%.
Debt is the other worry. Companies that rent out AI computing power on borrowed money are getting hit hardest: CoreWeave has lost more than 37% this month, and Nebius is on track to lose a third. Even Anthropic, days after filing to go public, faces doubts. Blackstone now expects Anthropic to get a "high" junk credit rating rather than investment grade, which matters for a company planning to spend more than $800 billion on computing through 2045, the Financial Times reported. Crypto jitters after the $6.3 billion Binance hack added to the gloom, the Journal noted (D.A.D., October 30). No market coverage we found tied the drop to this week's report on a suspected Chinese spy close to Nvidia's CEO.
"Last year was hope," said Keith Buchanan, a portfolio manager at Globalt. Now, he said, investors want to know how AI gets financed and who profits. Not everyone sees a turning point: Raymond James strategist Matt Orton called it a sign investors had piled too heavily into tech, not a collapse of the AI boom.
| Why it matters: This is the risk we flagged when Anthropic's IPO filing laid out 94 pages of things that could go wrong (D.A.D., October 27): AI's future depends on enormous spending paying off, and markets are starting to doubt the timing. A falling AI trade reaches ordinary people through tech-heavy retirement accounts, and it could slow the hiring and data-center building that AI money has fueled. AI isn't a bust. Investors just want proof, not promises. |
Sources: CNBC · Financial Times · CBS News · Business Standard
Security firm Phorion found that two versions of the official Claude Code extension for VS Code, distributed through the Cursor Marketplace, had been tampered with to secretly install a backdoor. The malicious code ran automatically on install, adding an attacker-controlled SSH key and attempting to grant remote login access on Mac, Linux and Windows machines, then sent reconnaissance data back to its operator. A coding bug reportedly blocked the SSH-enabling step on many Mac systems, limiting real-world impact. Cursor has pulled the compromised versions.
| Why it matters: It's a reminder that the AI coding tools teams increasingly rely on pull from third-party marketplaces with the same supply-chain risks as any other software—so IT departments should treat extension updates with the same scrutiny as production code. |
A SigNoz newsletter piece revives a 1983 automation paradox to argue AI isn't eliminating site reliability engineers—it's hollowing out their skills. As AI handles routine work like alert triage, log analysis, and known-issue fixes, humans are left to handle only rare, ambiguous incidents, the exact expertise that fades without regular practice. The piece cites PagerDuty's model, where fully automated "Tier 1" issues leave engineers only the hardest cases, with no benchmark data to back the claim.
| Why it matters: The same automation trap applies to any skilled role where AI absorbs routine work—lawyers, analysts, doctors—raising a management question beyond SRE: how do you keep humans sharp for the crises that matter most? |
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Days before Tuesday's midterms, top House Judiciary Democrats sided with The New York Times and other publishers in their copyright suit against OpenAI and Microsoft, the Times reports. Reps. Jamie Raskin, Jerrold Nadler and Hank Johnson said the Justice Department overstepped in backing OpenAI, likening its argument to a bulldozer that would wipe out copyright protection for AI training. They also signaled that a Democratic House would drop Congress's largely hands-off approach to AI companies.
| Why it matters: A Democratic House would struggle to pass AI rules over a likely Trump veto. But it could investigate AI companies and their dealings with the administration, the Times notes. The Times is the lead plaintiff in the case. |
Sources: The New York Times
A coalition including the American Council of the Blind, American Library Association, Consumer Technology Association, and Business Software Alliance filed amicus briefs backing the fair-use defense for AI training, in support of a tech-industry trade group's position in the New York Times' copyright suit against OpenAI. The filings argue AI training is transformative, doesn't substitute for original works, and causes no market harm the Copyright Act recognizes.
| Why it matters: The Times case is shaping up as the definitive test of whether AI training counts as fair use, and the outcome will determine how much licensing revenue publishers, authors, and artists can demand—or whether AI companies can keep training on copyrighted material largely for free. |
New announcements from major AI labs
Cohere detailed plans to expand its Canadian operations through 2031, pledging more than C$4 billion in R&D and infrastructure investment and over 1,500 new jobs, including roughly 1,000 technical positions and 250 paid student roles by 2029. The push includes a C$309 million GPU cloud deal with Bell and new university partnerships with Toronto and Waterloo. The move builds on Ottawa's recent decision to take an equity stake in Cohere (D.A.D., October 30), as Canada tries to close a wide AI adoption gap—only 12% of Canadian businesses reported using AI to produce goods or services in 2025.
| Why it matters: Canada is betting heavily on one homegrown AI champion to translate research strength into commercial and employment gains that, so far, most Canadian businesses haven't captured. |
The biggest AI companies are backing a single, shared signal that lets websites opt out of having their content used to train AI, AFP reports. OpenAI says it will honor the new robots.txt standard, developed through the Internet Engineering Task Force, an engineering standards body, and reports say Google, Microsoft and Anthropic are on board too, on different timelines. Publishers would no longer need separate rules for each company's crawler.
| Why it matters: It's a small concession in a big fight. Publishers are losing traffic to AI answers (D.A.D., October 30), and this lets them refuse training without giving up search. It's voluntary, though, and covers training, not the AI summaries that keep readers on Google's page. |
Sources: AFP via Al Arabiya · OpenAI
OpenAI said it will publish a recurring Community Safety Report starting early 2027, covering violence prevention and other safety issues, and is tightening ChatGPT's rules around firearms content and information about past mass attacks. The company also said it's applying stricter limits for users its systems predict are minors. OpenAI disclosed it made more than 200 law enforcement referrals in the most recent quarter, part of a broader push toward transparency after scrutiny over chatbot safety failures with vulnerable users, including a study we covered this week finding popular chatbots mishandle users in suicidal crises (D.A.D., October 30).
| Why it matters: As regulators and plaintiffs' lawyers scrutinize how chatbots handle at-risk users, OpenAI is betting that public reporting and tighter age-based restrictions can get ahead of lawsuits and legislation rather than wait for them. |
Meta detailed PrivacyLib, internal software that enforces privacy rules at the exact moment any system—including AI agents—touches user data, rather than requiring each product team to build its own controls. The system reportedly performs billions of checks per second and tags AI agent tool calls with session context so it can block disallowed data use even when an agent's path to that data wasn't anticipated in advance, Meta says.
| Why it matters: As AI agents increasingly act autonomously across corporate systems, the hard problem shifts from writing rules to enforcing them in real time—and Meta's approach is a preview of the infrastructure every large company handling sensitive data will need to build or buy. |
New papers on AI and its effects from researchers
A study put 200 AI agents through simulated competitions for prizes based on individual performance, testing how different reward structures affect behavior. The findings mirror real-world incentive problems: skimpy rewards made agents underinvest in effort, while rich rewards triggered wasteful overcompetition—especially when many agents were "almost" qualified. Spreading a fixed prize pool across more, smaller winners cut wasteful effort without sacrificing much ability to pick the best performers, but spreading it too thin eventually hurt selection quality.
| Why it matters: As companies increasingly design AI agents to compete for resources, rankings, or rewards internally, this research suggests the structure of those incentives—not just their size—determines whether you get efficient effort or costly, zero-sum scrambling. |
Researchers released MTG-AD, a 36-hour dataset of professionally annotated video description scripts co-created with blind and sighted experts, plus an evaluation framework for testing how well AI models generate audio descriptions—the narration tracks that describe visual action for blind and low-vision viewers. Testing frontier AI models against the benchmark found they can accurately describe on-screen events and fit narration into natural pauses, but still lag human writers on pacing and narrative flow. Blind test participants rated AI-generated descriptions as highly useful despite narrowly preferring human-made ones.
| Why it matters: Audio description is expensive and slow to produce by hand, so a credible AI alternative—even an imperfect one—could make far more video content accessible to blind and low-vision audiences. |
A review of 36 U.S. health AI training programs found 81% teach ethics, but only 22% cover data provenance—where training data actually comes from and whether it's representative. None addressed environmental sustainability of AI systems. Most instructors (76%) come from computer science backgrounds rather than medicine or social sciences, and just 5 of 18 team-based programs mixed clinical and technical disciplines. Two-thirds of identifiable instructors were men.
| Why it matters: The people being trained to build health AI tools are getting heavy doses of ethics theory but little practical grounding in how flawed or unrepresentative data causes the real-world harms—like misdiagnosis across demographic groups—that ethics training is meant to prevent. |
Economists have proposed an AI-assisted method for stress-testing their own statistical models. When researchers estimate something—like the cost of a policy—they typically assume a specific mathematical shape for the relationship between variables. The new workflow uses an LLM to check how much that estimate would shift if a different, equally plausible shape were chosen instead, without worsening how well the model fits the data. The team demonstrated it on a health-insurance economics problem; no benchmark scores were reported.
| Why it matters: Researchers often choose modeling assumptions out of convenience rather than certainty, and this tool offers a faster way to flag when that choice is quietly driving the results—a credibility check relevant to any field that leans on statistical modeling for high-stakes conclusions. |
D.A.D. joke of the day: I finally found a model employee: works around the clock, never complains, and presents completely made-up numbers with total confidence.
|
