AI News, Nov 16: Washington Starts Writing the Chip Rules
The Big Story: Commerce starts writing the rules the Nvidia scandal exposed
The Bureau of Industry and Security issued an interim final rule on remote access to advanced chips and servers, closing the gap where restrictions on American hardware could be sidestepped by renting time on it in a third country. Reaching a controlled chip over a network is now treated like putting one in a crate.
The most revealing sentence in the release is not about chips. BIS says the rule was written “by and at the direction of the Department of Treasury and the National Security Council.” An export-control agency announcing that it did not author its own export control tells you where chip policy now lives.
It landed the morning after Treasury Secretary Scott Bessent went on Fox News and said Nvidia’s board should recruit its next chief executive from outside the company, with other executives facing export-control scrutiny. He asked whether Nvidia had been “pressing these seemingly illicit sales to goose their revenues and their earnings,” then reached for a banking comparison: “If this were a bank and they hadn’t complied like this, the OCC would be in there right now.”
Three weeks ago this was a personnel story about one founder. It is now rulemaking, and rules outlast whoever ends up running Nvidia.
CVS put agents into production; two in five companies can’t reconstruct what an agent did
OpenAI shipped computer-use agents for healthcare that drive existing clinical software by “reading screens, clicking, typing, and moving between applications.” Abridge, Ambience, Cigna Healthcare and CVS Health are named as deploying them on referrals and on moving documents between systems, with each organization choosing which applications an agent may touch and which steps need human review.
Palo Alto Networks published a survey of 2,800 security practitioners across ten countries the same day. Seventy-three percent say agents can already modify code, pipeline configuration or production environments. Forty percent cannot reconstruct what an agent did, rising to 46% in financial services, and 53% log agent activity under a developer’s identity. Ninety-nine percent found ungoverned AI in their environment last year; 9% can discover and classify it automatically.
Google Cloud, also Monday, found that only 30% of healthcare organizations have scaled AI past pilots, 13 points behind the all-industry benchmark. The constraint is not capability. It is that the audit log says a developer did it.
Today’s Top Stories
Google found an exploit pipeline with no human in it
GTIG’s quarterly report documents a framework called exploit_agent, found on an exposed staging server and wired into several frontier and locally hosted models. It continuously compares widely used web application code before and after each patch announcement, drafts an exploit when it spots a flaw, and fires it, with no human in the loop. Google’s previous quarterly report had said it observed no fully autonomous attack pipelines deployed in the wild. The same report covers a threat actor using Cairn, a multi-agent pentest framework, during an intrusion in the US aerospace sector, and a group editing Claude Code’s settings inside victim repositories so that launching the coding assistant pulled down malware. Patch Tuesday now has a machine reading it.
Alibaba ships Qwen4-Max with the attack data taken out
Qwen4-Max is a sparse mixture of experts with a 2.4-trillion-parameter backbone, 71 billion activated per token and a million-token context. Alibaba scores it 52.69 on its own Work360 agentic benchmark, up from 47.74, with average cost per task falling from $4.43 to $1.67. The detail worth keeping is the training mix: Alibaba says cybersecurity data was filtered by task purpose, defensive material kept and data for “offensive tasks aimed at enhancing attack capabilities” removed. That published the same morning Google reported finding an autonomous exploit pipeline in the wild. Weights are not open, smaller Qwen4 models are promised as open source, and the Hacker News thread is largely people saying they will wait for those.
Anthropic’s IPO gets a ticker and a date
The listing is 825 million Class A shares at $70 to $78, pricing the evening of Nov 18 and opening Thursday Nov 19 on the Nasdaq as ANT. At that range it raises more than $64 billion at a $1.405 trillion valuation, with up to 20% of the offering targeted at retail investors. A retail allocation that size on the largest offering ever is a bet that the public wants to own the AI trade directly rather than through Nvidia.
The IMF counts more than $1 trillion in circular AI financing
Kristalina Georgieva told CNBC the Fund now puts circular financing in AI at a little over $1 trillion, roughly two-thirds of it debt-funded. “We have a world economy that is standing on one leg on the AI leg. If this leg wobbles, that can lead to financial stability troubles,” she said, while calling the total “not a huge size, but it is a rapid increase.” UBS analyst Tim Arcuri estimates a pullback in Nvidia’s financing could remove $50 billion or more from the system. Two-thirds debt-funded is the number that matters: equity losses stay with investors, debt losses travel.
Microsoft blames the 365 outage on a July patch
The root-cause write-up names three compounding factors: a latent defect in a July patch for Microsoft 365 networking gateway servers that load testing missed because it focused on “weaker hardware SKUs with fewer cores”; a load surge after Teams failed over onto the primary gateway, which also fronts Exchange Online, some fifteen hours earlier; and nearly 14% of gateway capacity already offline for an unrelated reason. More than 80% of servers were unresponsive at the low point, reached in under two hours. Because the bug was months old, rollback, flighting and feature flags could not revert it. Deployment safety tooling assumes the bad change is the recent one.
Quick Hits
- Robotaxis: Zoox vehicles now get matched to UberX and Uber Comfort requests in Las Vegas, Zoox’s first third-party ride-hailing deal, under a Nevada permit covering up to 1,000 robotaxis in Clark County.
- Security funding: Corma raised $125 million at a $1 billion valuation led by NEA, about a year after founding, for defensive security agents that work through a company’s existing tools.
- Venture: Plural closed €1 billion ($1.1 billion) for European moonshots in energy, AI infrastructure, defense and health, two-thirds of it a third early-stage fund.
- CRM: Freshworks and Salesforce alumni raised $53 million for Dreamteam, an AI-native CRM aimed squarely at Salesforce and HubSpot.
- Health: Apple launched a study app built with OpenEvidence, targeting up to 200,000 US participants over more than 12 months, with answers grounded in the clinical literature.
- Hacker News: “All AI-generated code should be Open Source” drew 1,200 points and 640-plus comments, and the best-received rebuttal was that if AI output isn’t copyrightable, a licence on it does nothing.
- Energy: Bedrock Energy raised $24 million for AI-directed geothermal drilling, and battery maker Our Next Energy is going public via SPAC on the strength of data-center grid strain.
- Legal: HSF Kramer hired six legal engineers out of Norm.AI, a sign the law-firm AI build-out has become a hiring market.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
