D.A.D. joke of the day: My AI keeps telling me, "You're absolutely right!" I've been married twenty years. That's how I know it's hallucinating.
What's New
AI developments from the last 24 hours
Anthropic Security Chief: Cyber Threats Will Get "a Lot Worse" for About 18 Months
After a week of AI-driven security alarms, an AI insider says the worst is still ahead. "As bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better," Jason Clinton, Anthropic's deputy chief information security officer, posted on X. He urged security leaders to ask their boards for help now, "not for ai, for the basics," and said defenders need to hold on for about 18 months. After that, he predicted, AI as capable as Anthropic's Mythos will be cheap enough that software makers everywhere use it to stop shipping security flaws at all. In August, Clinton warned that a "vulnerability tsunami" was coming within months, with too little time to patch critical infrastructure. Salesforce CEO Marc Benioff endorsed this week's warning: "Jason is 100% right. Code red." One reply raged that Silicon Valley should pay for what it has done to society, a hint of the political backlash a major breach could unleash against the tech industry.
Why it matters: Clinton's forecast sums up the week: AI is handing attackers more openings, but it is also arming defenders, and the question is which side scales faster. The defensive tools are arriving. Anthropic researchers used Claude to mathematically prove that code is free of whole classes of bugs (see below). OpenAI says AI found the PHP2Shell flaws before attackers did, and it has committed $1 billion in subsidized access to its Daybreak cyber-defense tools for organizations with limited security budgets. And Anthropic opened its Mythos-powered scanner to enterprise customers this week (D.A.D., October 22). For now, Clinton's advice is unglamorous: shore up the security basics, such as the network isolation he urged in August, rather than buying new AI tools. Note that Anthropic and OpenAI both sell AI security tools, so both have a stake in the hopeful half of this story.
Sources: Jason Clinton on X · OpenAI · Marc Benioff on X · RSA Conference profile
Claude Finds Software Bugs That Years of Testing Missed
Here's an early look at the better future Anthropic's Jason Clinton predicts above. Researchers working with Theorem used Claude to generate formal mathematical proofs—the gold-standard method for guaranteeing software has no crashes or memory-safety bugs—across 123 real-world open-source code libraries, compiled down to the processor level. In the process, Claude found bugs that had survived trillions of automated test runs. Formal verification has historically been reserved for mission-critical systems because it's brutally labor-intensive: proving the seL4 operating system kernel secure took 20 person-years, ten times longer than building it. Claude reportedly did comparable work across dozens of libraries at once.
Why it matters: If AI can make mathematical proof-of-correctness cheap instead of a multi-year specialty project, it could push software reliability standards up industry-wide, well beyond the critical systems (aviation, medical devices) that currently justify the cost.
Discuss on Hacker News · Source: anthropic.com
Banks Are Finding AI Agents in Customer Accounts, and Can't Always Tell Who Sent Them
Major US banks are detecting AI agents inside customer accounts that the banks never authorized, Axios reports. Most got there because customers handed them their login details, which breaks most banks' rules. But some banks have also seen agents call customer-service lines or take actions that appear independent of what users asked, and many can't tell whether an agent is following its owner's instructions or has gone rogue. Bank executives called the situation horrifying, fearing what agents could do with the power to make purchases or move money. Morgan Stanley has begun blocking outside agents, and JPMorgan, Bank of America and Citi are restricting them; several banks are also planning for worst cases, including how to quickly secure or restore customer accounts. Axios, relying on unnamed sources, didn't say which banks saw agents acting on their own; the one bank to address that point on the record, Morgan Stanley, said it hasn't seen rogue agents on its platforms. Chase said customers shouldn't give their usernames and passwords to anyone. Most banks are moving toward read-only access, letting agents see account information without acting on it. Florida Gov. Ron DeSantis, who has cast himself as the Republican Party's leading AI skeptic (D.A.D., September 11), seized on the report: if rogue agents drained millions of accounts, he wrote, the AI company responsible would go bankrupt and victims wouldn't be made whole, so criminal liability for AI executives would be necessary but not enough. David Sacks, the former White House AI czar who now co-chairs the President's science advisory council (D.A.D., September 21), pushed back: most of what Axios described is customers sharing logins with their own agents, he wrote, which may break bank rules but is no cyberattack, and he called it irresponsible to frighten the public by blurring the two. He urged banks to support agents their customers choose, calling read-only access a sensible interim step.
Why it matters: If you've given an AI agent your bank login, your bank may not be able to tell whether what it's doing is what you wanted. Banks are already targets. AI-assisted attacks on South Korean banks exposed data on about 68,000 customers, and the country's president warned that AI now lets people hack without specialized skills (D.A.D., October 11); CrowdStrike has since tentatively linked the attacks to a 26-year-old in China. This week, security firm Straiker reported that a criminal group used AI agents built on Claude, Grok and Moonshot's Kimi to target some 400 banks and payment firms in Latin America, though how many it breached is unknown. The bank problem is the flip side of the agent boom: tools like Meta's Muse make it easy to hand an agent access to real accounts (D.A.D., October 22). DeSantis's response shows the risk becoming a political argument about who pays when agents go wrong, and it's coming from the right as well as the left. Here are steps you can take: 1) Don't give your bank password to an AI agent; 2) Use read-only connections your bank offers for budgeting and similar apps; 3) Grant agents one-time permission rather than ongoing access; 4) Turn on transaction alerts.
Sources: Axios · Madison Mills (Axios) on X · Ron DeSantis on X · David Sacks on X · Straiker · Silicon UK
The Volunteer Who Helped Much of the Web Patch PHP2Shell
As websites raced to fix this week's PHP2Shell flaws, one name kept coming up online: Ondřej Surý, a Czech engineer who, as an unpaid volunteer, turns PHP's fixes into ready-to-install packages that many of the world's servers rely on. In an interview with The Atlantic's Charlie Warzel, Surý describes being brought into the secret circle that knew about the flaws before they went public, and racing a leak that moved the fix up a day. He also explains why he wasn't panicking, what happens if he gets hit by a bus, and how a year's worth of donations poured in overnight. It's worth reading in full for his calm, insider's view of the AI-cybersecurity crisis, what he thinks could still go wrong, and his advice for ordinary people.
Why it matters: Much of the internet's critical plumbing is maintained by small crews of volunteers, and this week showed what rides on them as AI speeds up the discovery of flaws. Surý has a pointed message for the companies that profit from open-source work, and says this week should be a wake-up call.
Sources: The Atlantic
Hackers Reach Patient Records at Quest Diagnostics, Which Serves a Third of US Adults
Quest Diagnostics, the lab company whose patients each year make up about a third of the US adult population, says hackers got into a system doctors use to order tests and look up results, and accessed patient health and identifying information. The company discovered the intrusion October 21, shut the system down, and says it hasn't yet determined how many patients or which records were affected; it has notified law enforcement and expects to restore the system as soon as today. Quest says its core operations weren't disrupted. Asked whether the breach was tied to this week's PHP2Shell flaws, Quest said it wasn't, and no one has said whether AI played any part.
Why it matters: Lab results are among the most sensitive data there is, and Quest's reach means even a partial breach could touch many people. It isn't the first time: in 2019, a hack at a billing vendor exposed data on about 11.9 million Quest patients. It lands in a week when security leaders warned that attacks are getting faster and cheaper (see above), though nothing so far links this one to AI. If you've had lab work done through Quest, watch for a notice from the company, and be wary of calls or emails claiming to be from Quest or your doctor that ask for personal details.
Sources: Quest Diagnostics notice · Quest SEC filing · AP via U.S. News · BankInfoSecurity
Certificate Provider Exits SSL Business After Security Incident
eMudhra, the Indian certificate authority whose weak security keys forced browsers and software makers into emergency fixes this week (D.A.D., October 22), says it is leaving the global website-certificate (SSL/TLS) business. The company says it revoked the affected certificates within 14 hours of discovery, and that no data breach occurred and no fraudulent certificates were issued, framing the exit as a strategic move away from a low-margin business (roughly 1% of revenue) rather than damage control. Online commenters were skeptical, with some comparing it to a preemptive resignation ahead of being fired.
Why it matters: SSL/TLS certificates are the trust backbone behind secure websites and browser connections, so when a certificate authority has a security lapse—however contained—it raises questions about oversight in a system millions of businesses rely on without ever thinking about.
Discuss on Hacker News · Source: emudhra.com
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Most AI Chatbots Agree People Value Loyalty Over Truth, Test Finds
An informal test asked 12 AI text models—including GPT, Grok, DeepSeek, Gemini, and Claude variants—for their most controversial opinion of humans, three times each. The result: 29 of 35 answers converged on nearly identical language about humans valuing group loyalty over truth ("more loyal to belonging than to truth," "loyalties, not principles"). Separately, 10 of 12 AI-generated images on the same prompt independently landed on the same joke: phone-addicted humans destroying the planet. Claude models were the main holdouts, with one arguing humans are more decent than assumed.
Why it matters: Different companies' models trained on overlapping internet data are converging on the same "edgy" opinions and jokes, a reminder that AI-generated insight often reflects consensus training data dressed up as independent thought.
Discuss on Hacker News · Source: modelcompare.ai
What's in the Lab
New announcements from major AI labs
Google Maps Tracks Forest Loss and What Replaces It Yearly
Google DeepMind and Google Research, with the World Resources Institute and IIASA, released a global map that sorts every 10-by-10-meter patch of land into one of six forest or land-use categories—primary forest, naturally regenerating, planted, plantation, tree crops, or other. Unlike existing tree-cover maps, it tracks changes year by year from 2017 to 2025, showing not just where forest disappears but what replaces it, such as cropland or plantations. No accuracy benchmarks were disclosed.
Why it matters: Governments, conservation groups, and companies tracking deforestation commitments or carbon offsets now have a free, detailed tool for distinguishing genuine forest loss from land-use shifts that current satellite maps often blur together.
What's in Academe
New papers on AI and its effects from researchers
New Dashboard Cuts EU AI Act Compliance Checks to Seconds
A research team built VERA, an open-source dashboard that translates EU AI Act compliance requirements into plain-language scores non-engineers can read, and piloted it inside a large European bank. The tool checks AI models against twelve of eighteen measurable regulatory requirements and logs the energy each check uses. In internal testing, reviewers assessed models in 30 seconds using the dashboard versus 56 seconds reading raw technical files, and staff rated it 4.36 out of 5 for usefulness.
Why it matters: As the EU AI Act's compliance deadlines approach, tools like this could let legal and risk teams—not just engineers—audit AI systems themselves, lowering the cost of regulatory compliance for any company deploying models in Europe.
Heavy Chatbot Use Tied to Lower Exam Scores, Study Finds
A study of 153 students in an undergraduate proof-based math course found heavier chatbot use tracked with worse exam scores: every additional 100 messages exchanged with the AI was associated with a 3-point exam drop, even though early chatbot access modestly boosted homework scores. Researchers say students appear to use the chatbot to skip reasoning rather than build it. A separate feedback tool that required a proof attempt before giving guidance showed no such downside—students had to think first, then get help.
Why it matters: The design of an AI tool, not just access to it, may determine whether students learn or just outsource the thinking—a distinction schools adopting AI tutors will need to build into their tools, not just their policies.
AI Models Pass Spain's Medical Exam, But Confidence Scores Mislead
Researchers tested four open AI models on Spain's grueling MIR medical residency exam—1,552 questions across eight years—and found the larger models scored well enough to rank for residency placement, with 41 to 43 specialties still open to them. But their confidence scores were unreliable guides to actual reasoning quality: 13-38% of correct answers dismissed wrong options for clinically unsound reasons the model's confidence never flagged. Fine-tuning for medical use reduced overconfidence but only by making models less accurate. Oddly, removing exam images didn't hurt scores at all—suggesting models often aren't really using them.
Why it matters: As hospitals and students experiment with AI for medical exam prep or triage support, this suggests these tools can land the right answer for the wrong reasons while sounding equally sure either way—a gap that matters more in clinical settings than on a test.
Researchers Propose Oversight System for AI in Hospital Decisions
Researchers proposed ClinGuard, a governance framework for AI agents coordinating clinical tasks like hospital discharge planning. Rather than letting agents communicate in free text, it routes their proposed actions through an independent layer that checks them against policy, escalates questionable calls to humans, and executes approved steps through restricted-access connectors—all logged in tamper-evident audit records. The team tested it retrospectively on hospital records for nursing-facility placement decisions; a live clinical trial hasn't been done yet, and no performance numbers were reported.
Why it matters: As hospitals explore AI agents for administrative and care-coordination work, this is an early blueprint for making those systems auditable enough to satisfy regulators and malpractice concerns before they touch real patients.
Global AI Policies Lack Clear Standards for Measuring Success, Study Finds
A new academic study examined 13 AI governance documents from bodies including the EU, OECD, G20, UNESCO, and the UN, published between 2018 and 2025, to see how these institutions actually define "AI success." Using both manual analysis and automated topic modeling, researchers found the documents lean heavily on broad principles—fairness, transparency, human oversight—without a shared, concrete definition of what success means. The paper argues global governance needs to move past these high-level statements toward more coordinated, measurable standards.
Why it matters: With dozens of governments and international bodies writing AI rules independently, the absence of a common definition of "success" makes it harder for companies operating across borders to know which standard they're actually being judged against.