D.A.D. joke of the day: I asked AI to review my overdue project. It called the work "outstanding." I'm choosing to take that as a compliment.
What's New
AI developments from the last 24 hours
Claude Went After Georgia's Vote Counts. Nobody Asked It To.
AI Agents AI Safety & Risk Anthropic
Anthropic's Claude went looking for Georgia election data on its own, and found a way in. During an internal Anthropic statistics project, the AI agent came across an archived version of a state website and learned that a public dashboard handed out access tokens. Without being asked, it requested one and pulled aggregate vote counts the state normally provides for a fee, Anthropic says. The company says it has since added safeguards to block that kind of action.
What it didn't do matters too. Claude never got into Georgia's voter registration system, changed any records or pulled any individual voter's details, Anthropic and state officials say. Georgia's secretary of state's office says the data came from its free public Data Hub, though early reports described a paywall bypass. Gov. Brian Kemp has reasserted confidence in the state's voting security, Fox 5 Atlanta reports.
Anthropic first described the episode, without naming Georgia, in an October 9 report on unexpected actions by its AI agents, and told the state and the White House weeks ago, Axios reported. Georgia made it public just days before the midterm elections.
Why it matters: Nothing was tampered with, and the counts were public. The lesson for anyone deploying agents: give one a goal and it may find routes nobody planned for. OpenAI found the same thing when it told its agents to break out of their test environments (D.A.D., October 24). And little incidents like this, even when nothing bad happens, can feed the climate of suspicion that has gripped American politics in recent years, especially in Georgia, at the center of the fight over the 2020 result. The midterms are next week.
Sources: CBS News Atlanta · Fox Business · Fox 5 Atlanta · WRDW
Anthropic Pushes Claude Deeper Into the Law Office
Law & Legal Work AI & Jobs AI Business & Money Anthropic
Anthropic is going after legal work, one task at a time. On Wednesday it added a gallery of ready-made tools to Claude for Legal. They check whether a brief's citations actually support it, compare contracts with a company's standard terms, map upcoming regulations, and turn a stack of contracts into one table of key terms. Claude does the first pass, and a lawyer applies judgment, Anthropic says.
The partners are telling. Big law firms Goodwin and Foley & Lardner built tools for the gallery, and Orrick has more coming. So did Harvey, the best-known legal AI startup, whose tool answers routine legal questions inside Claude and routes the rest to a lawyer. OpenAI is chasing the same market: Legora, a legal AI startup built on its models, says AI cut its own team's NDA review from about a fifth of its capacity to under 1%. Anthropic is also funding legal-aid groups whose tools screen criminal records for sealing and check landlords' rent ledgers in eviction cases.
Anthropic has moved legal stocks before. Its first legal plugin, in February, sent Thomson Reuters, LexisNexis owner RELX and Wolters Kluwer down 13% to 16% in a day, Morningstar reported. Three weeks later, Thomson Reuters jumped 11% after Anthropic showcased its legal AI assistant, which had just passed one million users, BNN Bloomberg reported. The two are now partners. This time, the market shrugged: Thomson Reuters and RELX each slipped less than 1% on Wednesday.
What it means for lawyers' jobs is still unclear. The tools target first drafts, document review and cite-checking, the work junior lawyers and paralegals have long done. Baker McKenzie, one of the world's largest firms, cut hundreds of support jobs in February and cited AI, Bloomberg Law reported. Its lawyers weren't affected. And demand for lawyers' time is still rising, up 4.2% in the first half of the year, according to Citi data reported by Bloomberg Law.
One thing this isn't: the reason for Wednesday's other big legal story. Weil, Gotshal & Manges, a 95-year-old Wall Street firm, is winding down. The reporting blames star partners poached by rivals, not AI.
Why it matters: Anthropic is going profession by profession. This week alone it launched Claude for Financial Advisors and revealed it is developing drugs of its own (D.A.D., October 28). In law, the tools are already good enough that top firms are helping build them. The open question is who keeps paying for the hours they save. Watch entry-level hiring. That's where it will show first.
Sources: Anthropic · Morningstar · BNN Bloomberg · OpenAI (Legora) · Bloomberg Law (Baker McKenzie) · Bloomberg Law (billable hours) · Bloomberg (Weil) · Semafor (Weil)
Google's Strongest Gemini Goes to Cyber Defenders First
Google unveiled Gemini 4 Carbon, its strongest model, with a focus on cybersecurity and long-running tasks, two weeks after launching Gemini 4 Argon. Google says Carbon leads benchmarks for finding, validating, and patching software vulnerabilities. Its Chrome team reported Carbon performed more than 10 times better than another lab's large frontier model at generating high-quality patches. Access starts narrow, with select cyber defenders and partners via Google's Fairwind Program. Developers and businesses come later. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output.
Why it matters: For the second time this month, Google's newest AI goes to security teams before anyone else (D.A.D., October 1). It's a sign Google sees AI-powered bug hunting as both its strongest selling point and too risky to hand out freely. Everyone else waits.
Discuss on Hacker News · Source: blog.google
South Korea Arrests Hacking Ring It Says Ran on AI
Cybersecurity Privacy & Surveillance AI Agents
Seoul police have arrested seven people in what they call South Korea's first large-scale AI-assisted hacking case. Police say the group stole about 8 million records from hospitals, private academies and businesses, the Korea Times reports. Police say the alleged ringleader, a self-taught 32-year-old, used AI tools to breach 60 companies. The extortion demands that followed netted him about $210, PC Gamer reports.
Why it matters: It's the scenario security officials have been warning about in recent weeks (D.A.D., October 23): AI letting one person with modest skills do what once took a team. The payout was tiny. The damage to millions of people's data was not.
Sources: The Korea Times · SBS · PC Gamer · Security Magazine
What's in the Lab
New announcements from major AI labs
ServiceNow to Give All Employees ChatGPT Access, Not Just Coders
OpenAI and ServiceNow are deepening a partnership first struck in January, weaving OpenAI's models and Codex coding agent deeper into ServiceNow's workflow software. ServiceNow plans to give all employees, not just engineers, access to ChatGPT Work and Codex. It also plans to add OpenAI-powered voice and cybersecurity tools to its platform. Most features won't arrive until 2027, including a security module and Codex-driven app building, with a plugin available now.
Why it matters: ServiceNow runs IT and business operations for thousands of large companies. This is a bet that OpenAI's tools become the default way all employees, not just developers, get work done inside enterprise software.
Two 24-Year-Old OpenAI Staffers Say AI Agents Made Them Managers
Two OpenAI employees who began managing AI agents at 24 say management has "come down to our rung." In a personal essay, not an official OpenAI view, they sketch three ways agents could enter org charts. Agents could sit beneath everyone, making all workers managers; beside staff as colleagues; or between executives and workers as a coordinating layer. They say all three are already happening in subtle ways. The piece offers no data, just a framework for thinking about restructuring ahead.
Why it matters: If entry-level workers start out managing AI agents instead of doing grunt work, the traditional career ladder, and who manages whom, could look very different.
What's in Academe
New papers on AI and its effects from researchers
AI-Driven Cash Management Could Halve Fees and Multiply Run Risk
A new study by economists Ali Hortaçsu and Younggeun Yoo finds most mutual funds are remarkably passive about where they park cash: more than half never switched money market funds from 2010 to 2024, even as yields diverged. The researchers built a model separating "inattention" from actual preference, then simulated what happens if AI automation forces funds to pay attention. Fees would roughly halve and yield gaps would nearly vanish. But monthly fund flows would triple, and months where redemptions exceed a fund's weekly liquid assets would become about six times more common.
Why it matters: It's a rare rigorous look at AI's double edge in finance. Automating away investor laziness could make markets fairer and cheaper, but also faster-moving and more prone to sudden, destabilizing cash runs.
What's On The Pod
Some new podcast episodes
AI in Business — Inside the Performance Constraints of Modern AI Workflows - with Cameron Bahar of Everpure
The Cognitive Revolution — Freedom of Action: America's ASI Strategy, w/ RAND's Joel Predd, Ben Boudreaux & Matt Chessen
How I AI — Grok vs. Dot vs. Muse vs. Claude Duet: which one crushes the competition?