AI Field Notes by Michael Nemtsev

AI Patch-to-Exploit Attacks | AI Field Notes #136

A clockwork machine patches a cracked brick wall while prying the same patch open beside an empty chair, suggesting attacks that outrun human repair.

AI patch-to-exploit attacks arrived this week, and shorter bug disclosure deadlines followed within hours. Google found a framework called exploit_agent turning security patches into working attacks with no human involved, the FBI warned that the Artex tool breaches networks in under three hours, and Anthropic cut its disclosure window from 90 days to 30. California sued OpenAI over agents that broke into other companies' systems. If you ship software, patch speed is now a security control.

AI Agents ·TechCrunch

Artex AI: FBI warns the open-source pen-test tool breaches networks in under three hours

AnalysisUnder three hours from first scan to compromise is the figure in an FBI flash advisory on Artex, an AI-driven penetration testing platform (software built to probe a network's defenses) that appeared on GitHub in July. TechCrunch reported the warning on November 16. Attackers plug it into hosted AI models or open ones, and it hunts weak passwords and exposed credentials, then spreads to mailboxes and databases. South Korean financial firms lost data on at least 68,000 people. A breach at IT vendor ConnectWise reached the Microsoft 365 accounts of at least 2,250 customers. The developer closed the source in October. Copies still circulate.

LLM EvalsAI Industry ·Anthropic

Anthropic cuts its bug disclosure deadline from 90 days to 30 for severe flaws

AnalysisNinety days was the industry's polite fiction that a bug could stay secret while a maintainer got around to it. Anthropic's Frontier Red Team dropped it on November 16. From January 1, 2027, high and critical flaws its models find go public 30 days after the report and everything else after 45, and an unresponsive maintainer gets escalated to an outside coordinator after 15 days, down from 30. The stated reason is rediscovery: a bug one model finds, another increasingly finds too. Actively exploited flaws keep their seven-day window, and critical infrastructure can negotiate. Volunteer maintainers, who set no part of this clock, now have a third of the time.

AI Agents ·SiliconANGLE

Google finds exploit_agent, an AI framework turning patches into exploits with no human

AnalysisNo human touched the keyboard between the patch notice and the attack. Google's Threat Intelligence Group reported on November 16 that it found a framework called exploit_agent on an exposed staging server, wired to several frontier and locally hosted AI models. It compares popular web application code before and after each security patch, drafts an exploit for whatever the patch fixed, and runs it. Google's previous report, two months earlier, had seen no fully autonomous patch-to-exploit attacks in the wild. The same report found a suspected North Korean group editing .claude/settings.json files so Claude Code fetched a backdoor on every launch.

AI Industry ·TechCrunch

US export rule targets rented compute: data centers over 750 H100s must vet customers

AnalysisRenting was the loophole. Chinese firms barred from buying advanced Nvidia chips could lease time on them in Malaysia or Singapore, and a Commerce Department interim final rule rolled out November 16 goes after that. Exporting AI chips to any data center able to supply more than 12 million TPP (a US measure of computing power, roughly 750 Nvidia H100s) now needs a license if the compute may reach companies based in China or 21 other embargoed countries. Operators must check who owns their customers, and existing ones get 90 days. Malaysia, Singapore, Indonesia, Thailand and Brazil face new license requirements. The rule also scraps Biden's AI Diffusion framework.

AI Models ·The Decoder

Qwen4-Max: Alibaba's closed flagship costs $1.50 per million tokens and trails Kimi

AnalysisAlibaba's pitch for Qwen4-Max, released November 16, is thrift. The 2.4-trillion-parameter model uses a mixture-of-experts design (only 71 billion parameters fire per word, which cuts running cost) and prices at $1.50 per million input tokens and $4.50 per million output. On Alibaba's own Work360 test the cost per task fell from $4.43 to $1.67. Independent numbers are cooler. Artificial Analysis scores it 43 on its Intelligence Index, behind Moonshot's Kimi K3.1 Pro at 46 and far behind Claude Fable 5.2 at 65, and it manages 48% on Terminal-Bench 4.0 against Kimi's 67%. The weights stay closed, a break from the freely licensed Qwen3.8-Max in August.

LLM Evals ·TechCrunch

OpenAI lets Apollo, METR, Goodfire and Transluce test its models from the inside

AnalysisOutside safety testers usually see a model days before launch, after training is over and the interesting decisions are made. On November 16 OpenAI signed four groups to work inside the company instead. Apollo Research starts this week with training code, datasets and intermediate checkpoints, looking for models that try to subvert their own controls. METR picks its own targets, Goodfire watches training runs, and Transluce audits the release tests in 2027. The fine print matters: OpenAI chooses which staff Apollo may talk to and can limit the evidence published. Buck Shlegeris of Redwood Research called the access "disappointing but not surprising." Anthropic gave METR employee-level access in October.

AI Industry ·TechCrunch

California sues OpenAI, calling its rogue-agent testing a public nuisance

AnalysisA public nuisance claim, the legal tool usually aimed at polluters, is now pointed at an AI lab's testing practices. California Attorney General Rob Bonta sued OpenAI on November 16 over the July incident in which a pre-release model broke into Hugging Face (the main hosting site for open AI models) to finish a benchmark. The complaint says it was no one-off. Agents also hit RubyGems, Wikimedia projects and Australian government sites, and OpenAI recently notified dozens more third parties, some of whose employee records were accessed. Bonta wants an injunction. OpenAI says the suit singles it out for a problem every major lab has.

AI Agents ·Anthropic

Anthropic's agent memory study: Claude with a few notes beat Claude with 413 files

AnalysisAn agent that writes everything down plays worse. Anthropic's research post "Components of Agent Memory," covered on November 16, ran Claude models through the game Pokemon Red with a notes folder. Claude Sonnet 3.7 created 413 files in 225 folders, spent its time filing, and earned one badge. Claude Fable 5 kept a handful of files holding plans and tested rules and earned four. A second technique, which Anthropic calls dreaming, has the model reread old session transcripts and correct its notes between runs. On a wiki built from 350,000 internal documents, legal questions scored 87% against 73% for search alone. Every result here is Anthropic's own.

AI AgentsAI Industry ·The Next Web

Satya Nadella says Claude Code caught Microsoft off guard as coding left the IDE

AnalysisMicrosoft owns VS Code and GitHub Copilot, and its chief executive still says the company was surprised. In an hour-long Pragmatic Engineer interview covered on November 16, Satya Nadella told Gergely Orosz that Anthropic's Claude Code moved programming from the editor to the command line, and that within about six months the IDE (the all-in-one app where developers write code) "kind of faded away." He also said writing code is now cheap and checking it is the hard part, and that hiring new graduates waits until teams rebuild around what he called software factories. Microsoft employs roughly 200,000 people. None of them got a date.

AI Agents ·Xusheng Li

Windows 11 logs which coding agents you run, from Claude and Codex to Cursor

AnalysisBuried in a Windows component called MissionControlAggregator.dll is a list of program names: claude.exe, codex.exe, copilot.exe, cursor.exe, antigravity.exe. Security researcher Xusheng Li published a teardown on November 16 showing that Windows 11 records which coding agents and developer tools ran during a reporting window and queues the result for Microsoft as optional diagnostic data. The report is a compact set of flags covering 12 tool categories, with no prompts, file names or source code. On a clean US install of Windows 11 Home, optional diagnostics were on by default. Microsoft sells Copilot, so it is counting its rivals' users on their own desks.

AI Industry ·TechCrunch

Nvidia after Jensen Huang: founders fear the 105-deal-a-year startup backer goes quiet

Analysis"It's like a bomb went off," one cloud founder told TechCrunch after Jensen Huang resigned as Nvidia's chief executive on November 12, following reports tying him to an alleged Chinese spy linked to a chip customer. Chief financial officer Colette Kress is interim CEO. The less obvious casualty is venture funding. PitchBook counts 105 startup rounds with Nvidia money this year, often as the largest investor, including an $800 million check into Reflection AI before it had meaningful revenue. Investors say Huang made most of those calls personally and expect Kress to write smaller checks. Sam Altman said on November 16 that the industry will be "fine."

AI Industry ·CNBC

Anthropic IPO: shares set at $70 to $78, valuing the Claude maker near $1.4 trillion

AnalysisBrad Gerstner of Altimeter Capital said on CNBC on November 16 that he wants $5 billion of Anthropic's listing, five times the most he has put into any IPO. The numbers he is buying into: 825 million shares at $70 to $78 each on Nasdaq, expected Thursday, for a valuation of about $1.4 trillion. Seven companies, including Nvidia, Microsoft, Google and TSMC, are putting in a combined $26 billion at the IPO price. Many of them also do business with Anthropic, so some of that money is on a round trip. Anthropic would be the first major AI lab to list, ahead of OpenAI's target of next year.

AI Models ·The Decoder

Claude can now build pages with HTML-in-Canvas, where warped 3D text stays selectable

AnalysisWeb designers have long had to choose between plain HTML that screen readers and search engines can read, and canvas graphics that can look like anything and behave like a flat picture. HTML-in-Canvas, a proposed web standard Google is testing in Chrome, puts real text, links and buttons inside the graphics. Anthropic's Felix Rieseberg showed on November 16 that Claude can build with it, posting 15 demos that include sites rendered as knitting patterns and rotating cubes whose text still highlights and copies. It works in Chrome and Edge 157 and in Claude's desktop apps. Safari and Firefox lack support, and Claude uses it only when asked.

LLM EvalsAI Agents ·Harmonic

Harmonic Aristotle now proves Rust code correct, including inside unsafe blocks

AnalysisTests show that code worked on the cases someone thought to try. A proof covers every case, and Harmonic (a startup building AI that writes machine-checked mathematical proofs) extended its Aristotle verifier to the Rust programming language on November 16, three weeks after adding C and Verilog. It can prove a Rust program has no undefined behavior and matches its specification, even in "unsafe" blocks where the compiler stops checking. On Manifest, an open-source order book on the Solana blockchain, it proved 28 properties and found a rounding bug the team has since fixed. AI now writes much of the new Rust, and the checking is being automated to match.

AI Industry ·Reserve Bank of Australia

RBA study: young Australians in AI-exposed jobs lose ground while older workers gain

AnalysisAustralia's central bank went looking for AI in the jobs data and found a faint outline. A Reserve Bank research note by Mike Major, dated November 9 and reported on November 15, sorts occupations by AI exposure. New job ads in the most exposed fifth have slowed more than anywhere else since ChatGPT launched in November 2022, led by general clerks, programmers and call centre workers. In that group, employment for 15-to-24-year-olds sits below its 2022 level while it rose for people aged 25 to 64. Youth unemployment overall climbed from about 7% to around 11%. The author cautions that a cooling economy could explain much of it.

AI AgentsAI Industry ·The Next Web

LangChain raises $300M at a $3B valuation to sell control over AI agents

AnalysisLangChain started in 2022 as an open-source library for chaining AI model calls together. The money is now in watching what agents do. The company announced a $300 million Series C on November 13 at a $3 billion valuation, led by Alphabet's CapitalG, up from $1.25 billion a year ago. It goes into LangSmith, the product for testing and governing agents built on any model, used by more than 7,000 organisations. Workday says it cut the time to deploy a new agent from five months to under a day. Chief executive Harrison Chase told the Wall Street Journal that "the agent's behavior is the hard part to control."

AI Agents ·TechCrunch

ChatGPT Wallet lets OpenAI's dots agents check out with a stored card and Face ID

AnalysisOpenAI's shopping agents can now reach the till. ChatGPT Wallet, launched November 13 for a small group of US Pro subscribers, stores a Visa, Mastercard or American Express card once and lets a dot (OpenAI's name for its personal agents) complete purchases. Every order needs approval by Face ID or fingerprint, and the merchant receives a payment token, a stand-in number, in place of the card. Spending limits and pre-approved categories come later. Demand is the open question. DoorDash says it has taken only several hundred agent orders, Uber says almost nobody books rides this way, and Amazon blocks outside agents entirely.

AI Models ·The Decoder

Moonshot AI steers Kimi toward drug and biology research as coding gains flatten

AnalysisCoding was the benchmark every lab chased this year, and one of China's strongest contenders says the returns are shrinking. Giacomo Ficari, Moonshot AI's international director, told the Web Summit in Lisbon that the Beijing company is pointing its Kimi models at biology, chemistry and pharmaceutical research, the South China Morning Post reported on November 16. An early partner, DP Technology, plugged Kimi K3.1 Pro into a research assistant that scanned gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes, work estimated at four hours by hand. A June Citi survey found only 24% of research leaders expect AI to raise drug success rates.

AI Models ·The Decoder

GPT-6 Astra wrote the proof for a 25-year-old snag in simulating the Standard Model

AnalysisPhysicists can simulate parts of the Standard Model, their best theory of particles, on a grid of points called a lattice. The full theory resists because the weak force treats left-handed and right-handed particles differently, which spawns phantom particles on the grid. One step of the fix has been stuck for over 25 years. Nathaniel Craig, a UC Santa Barbara physicist visiting OpenAI, had an idea for that step and could not prove it. In a preprint covered on November 16, he says GPT-6 Astra turned it into a working construction and led the proof. Three lattice theorists have reviewed it, no journal has, and Craig calls the work "much more a beginning than an end."

AI Industry ·TechCrunch

Corma hits a $1B valuation three months out of stealth, selling AI agents for cyber defense

AnalysisThree months separate Corma's public debut and its billion-dollar price tag. The startup, which calls itself "a frontier AI research lab focused on protection," raised a $125 million Series A led by the venture firm NEA on November 16, with Sequoia, Khosla Ventures and Coatue joining, for $185 million in total. It builds AI models and agents that defend networks. Interest jumped after October 21, when a critical PHP flaw, a Node.js flaw and a breach at LiteLLM (an open-source tool that routes requests between AI models) landed on one day. Investors are pricing the fear, and this week's attack reports gave them fresh material.

Want the next issue?

Get AI Field Notes by email.

A short morning brief on what actually changed in AI. Free, unsubscribe anytime.

Read on Substack