Google's Project Zero has replaced its 90-plus-30 disclosure policy with a fixed clock that reaches 30 days in January 2028, fix or no fix, because it says language models hav

AI Briefing · Daily brief

Project Zero's 30-day disclosure clock; Fed says 87% now rank AI the top risk

Monday, November 16, 2026 · 9 stories · 14 min read · Read on the web

Google's Project Zero has replaced its 90-plus-30 disclosure policy with a fixed clock that reaches 30 days in January 2028, fix or no fix, because it says language models have collapsed the time between patch and exploit. The Federal Reserve's stability report puts AI at the top of its risk survey for the first time, with $680 billion of computing debt now a third of investment-grade issuance. The Wall Street Journal reports Washington will name more than two dozen AI 'national champions' for export, days before Trump meets Xi in Shenzhen on Thursday with chip curbs on the table and a US–China AI incident hotline still not operational. The FAA now blames a router feature and its contractor for Friday's ground stop, with zero evidence of an AI link; GXO signed for up to 1,000 Apptronik humanoids; Fractile raised $720 million at $7 billion before shipping a chip; and Datadog, Check Point and Corelight published the operational detail behind two agent-era threats.

The one thing to know

Put a date field on your vulnerability tracker. From February 2 a Project Zero report becomes public 60 days after the vendor hears about it, and from January 2028 it is 30, whether or not a patch exists. If your SLA for operating systems, browsers and hypervisors is longer than that, the gap is now on a public schedule.

1 of 9

Project Zero scraps the 30-day soak period: bugs go public 90 days after reporting now, 60 from February 2, 30 from January 2028, fixed or not

Google's Project Zero published its 2026 disclosure policy on November 13, replacing the '90 + 30' rules in place since 2021 with 'Full X': details are released X days after the vendor is told, regardless of fix status, severity or variants. The team's stated reason is that LLM tooling lets attackers find bugs and turn patches into exploits almost immediately, so patch soak time is 'no longer meaningful', and that it now finds several vulnerabilities a week where it once found a few a year. Vendors objected that a hard clock would expose unfixed bugs and push them to fix Project Zero's reports ahead of worse ones; the phased schedule is the compromise. In-the-wild bugs stay on Google's seven-day policy.

Security: Measure your patch SLA for browsers, operating systems, hypervisors and common libraries against a 30-day total from vendor notification. Subscribe to Project Zero's Reporting Transparency list, which names the product within a week of the report, and map each entry against your inventory with its expected publication date.
Executives: Project Zero's policies became the industry norm in 2015 and 2021. Expect other researchers, AI bug-finding platforms and regulators to adopt the same clock, and ask vendors in the next review how fast they can ship a fix or a mitigation.
Developers: The mechanisms Project Zero says it expects vendors to build are feature flags, input filtering, authenticated alternate update channels and hotpatching. If you cannot turn a code path off in production without a release, start there.
ML engineers: The policy is a direct consequence of model-assisted vulnerability discovery scaling faster than fixing: the CSA notes only 97 of 1,596 AI-found open-source bugs reported under Project Glasswing had confirmed patches by May.
Beginners: Google's elite bug-hunting team will now publish details of a software flaw on a fixed timetable even if the maker has not fixed it, because AI has made attackers fast enough that waiting no longer protects anyone. The timetable shrinks to 30 days by 2028.

Source: Help Net Security · Cloud Security Alliance research note · Project Zero policy post · Full story

2 of 9

Fed Financial Stability Report: 87% of market contacts name AI the top risk, up from 50% in May; $680 billion of computing debt is nearly a third of investment-grade issuance

The Federal Reserve's semi-annual report, released Friday, is the first in which AI leads the risks cited by the New York Fed's market contacts: 87% named it, against 50% in May and 30% a year earlier, ahead of a broad asset correction at 52% and higher long rates at 48%. The report groups the concern into valuations, cyberattacks including 'damage caused by frontier AI models themselves', and financing, where it counts $680 billion of advanced-computing debt issued through October, more than $1.1 trillion outstanding, up to $3 trillion of debt, guarantees and leases on infrastructure not yet started, and close to $1 trillion of large-bank exposure to the ecosystem. It compares the cycle with the late-1990s telecom build-out and says it may soon exceed it. The survey closed before Huang's resignation and the midterms.

Finance: The $3 trillion of commitments not yet commenced is the figure to set against Anthropic's $901 billion and Nvidia's $500 billion-plus financing programme, both of which get a public airing this week. The Fed's worry is interconnection, not any single default: banks lending to developers, holding SPE paper and funding the private credit that lends to the same developers.
Executives: If 87% of surveyed contacts called AI the salient risk in October, the number after the Nvidia scandal is higher. Report AI-linked counterparties, including cloud vendors, as a concentration to the board.
Security: A central bank has listed model misbehaviour next to state cyberattacks as a financial-stability channel. Expect supervisors to ask banks about frontier-model incident plans, not only about vendors.
ML engineers: Nothing changes in your stack, but the capital that pays for your compute now sits in a regulator's risk list. Price cuts and capacity promises will be judged against balance sheets, not benchmarks.
Beginners: The Fed asks investors what could break the financial system; this time almost nine in ten said AI, because so much borrowing now funds data centres. The Fed says banks and markets are still sound but the web of loans needs close watching.

Source: Investing.com · AI Weekly (CNBC figures) · Federal Reserve report (PDF) · Full story

3 of 9 · developing

WSJ: Commerce will name more than two dozen AI 'national champions', including Nvidia, OpenAI, Anthropic, Microsoft and Amazon, for bundled export with federal financing and fast licences

The Wall Street Journal reported on Sunday, citing people familiar with the plans, that the Trump administration intends to designate more than two dozen US AI companies as national champions and that Commerce Secretary Howard Lutnick could announce the list within about a week, the same week Trump travels to the APEC summit in Shenzhen. The designation sits inside the American AI Exports Program created by a July 2025 executive order: packages of chips, models and cloud that qualify receive government advocacy, federal financing and faster export-licence reviews. Commerce had 78 applications by July; in October Beijing cleared DeepSeek and MiniMax to sell their models abroad. The timing could still change.

Executives: If you sell AI systems overseas, a designation is a subsidy; if you buy from a champion, expect the bundle to be pitched to your foreign subsidiaries with government backing. Either way the list will shape which US stacks allied governments steer procurement toward.
Finance: A champion list is Washington picking winners the week Nvidia reports and Anthropic prices. Price the designation into the named firms and the omission into their competitors, especially neoclouds and smaller model vendors.
Security: Bundled exports with licence conditions give Commerce more visibility than piecemeal sales, which is the administration's answer to the Megaspeed diversion case. Critics argue it creates more of exactly that risk.
Beginners: The US government plans to pick a team of AI companies whose products it will help sell to other countries, with loans and quicker export approval, to compete with Chinese AI that Beijing has just cleared for export.

Source: The Next Web, reporting the Wall Street Journal · Full story

4 of 9 · developing

The week that prices AI: Nvidia reports Tuesday at 1:20pm PT, Anthropic prices Wednesday on a reportedly six-times-covered book, Trump meets Xi Thursday with chip curbs and an unlit AI hotline on the agenda

Nvidia releases third-quarter results at about 1:20pm PT on Tuesday with a 2pm call, Colette Kress's first as interim CEO; consensus is roughly $108.8 billion of revenue and options imply a move of around 10%. Anthropic prices its 825 million share offering at $70 to $78 on Wednesday for Thursday trading as ANT; Bloomberg reported on November 12 that orders were about six times the shares available, and bankers have told investors existing backers get priority. Reuters reported on Monday that Thursday's Trump–Xi meeting in Shenzhen is expected to be dominated by the prospect of tighter US chip curbs after the Huang scandal, that an emergency AI-incident notification mechanism agreed in September is not yet operational, and that there will be no new progress on the Board of Trade initiative. Lutnick testifies to Senate Banking on December 10.

Finance: The order of events matters: Nvidia's guide and any change to its financing programme land the day before Anthropic prices, and Anthropic's placement partners include Nvidia. A weak Tuesday makes a top-of-range Wednesday harder; a strong one makes a raised range plausible.
Executives: Supply and pricing signals for 2027 compute arrive in 72 hours: Blackwell and Vera Rubin timelines from Nvidia, Anthropic's public capacity commitments, and whatever chip language the summit readout contains.
Security: The incident hotline is not live. A cross-border AI incident this winter still goes through ad hoc senior contacts; plan on it.
ML engineers: Watch Nvidia's margin line for what it implies about accelerator rental prices and the open-weight price war, and Anthropic's first public-company quarter for how pricing and capacity announcements will cluster from here.
Beginners: Three events in three days will set the tone for AI investing: the biggest chipmaker's results, the biggest-ever startup IPO, and a US–China summit about who may buy which chips.

Source: Nvidia investor notice · TradingKey (Bloomberg order-book report) · Reuters via Investing.com · Full story

5 of 9

FAA update: 12,462 cancellations Friday and about 6,700 Saturday; Duffy blames a broadcasting feature on new L3Harris-installed routers, and the FAA chief says 'zero evidence of anything AI related'

Transportation Secretary Sean Duffy said the ground stop began when a new Juniper router at an FAA site in El Dorado, Arkansas, installed by contractor L3Harris during a copper-to-fibre switch, was turned on and a router in Amarillo installed three weeks earlier 'started to flood our system as well'; officials believe a feature that broadcast data more widely than necessary caused the overload, and the FAA has disabled it on all 69 routers of that type. Both devices are being examined at an L3Harris lab with the FBI involved. The FAA counted 12,462 cancellations on Friday, cancellations passed 6,700 on Saturday, and about 1,570 flights were cut by Sunday morning, most at American and JetBlue. FAA Administrator Bryan Bedford said there was 'zero evidence of anything nefarious and zero evidence of anything AI related', and officials said the Microsoft outage the same week was unconnected. HPE, Juniper's owner, said the devices 'did not fail'.

Executives: The lesson is about change management on legacy networks, not AI: a new device with a default feature took down a national system during a planned migration. If your own copper-to-fibre or on-prem-to-cloud cutovers lack a rollback plan tested against production traffic, this is the case study to bring to the next risk committee.
Security: A configuration fault produced the signature of an attack. Make sure your incident playbooks can distinguish the two quickly, because the FBI was called in before the cause was known.
Beginners: The FAA now says Friday's nationwide stop was caused by a newly installed network router sending too much data, not by a cyberattack or AI. Flights were mostly back to normal by Sunday, and the contractor is being blamed.
Finance: Duffy wants $50 billion and Trump $30 billion for a replacement system on top of $12.5 billion already approved; the funding request goes to Congress this week. Air-traffic modernisation vendors just became a line item.

Source: Yahoo News live coverage · Full story

6 of 9

GXO signs a letter of intent for up to 1,000 Apptronik Apollo humanoids over four years, robots-as-a-service, KPI-gated, first commercial units in late 2027

GXO Logistics, which calls itself the largest pure-play contract logistics company, agreed on November 12 to take up to 1,000 Apollo humanoid robots from Austin-based Apptronik across its global network over four years. Expanded trials start in 2027 with first commercial deployments in North America late that year; the rollout is stepped and each use case must meet jointly agreed performance, safety and operational KPIs before it scales. GXO pays per robot as a service rather than buying a fleet. The task list is picking, case and tote movement, packing and sortation, tested on more than 50 logistics tasks at Apptronik's 90,000 square foot Robot Park. GXO has run over 45 humanoid pilots, is expanding Agility's Digit to two more sites, and said in August it had not yet achieved ROI on humanoids.

Executives: Copy the shape before the number: no purchase, a letter of intent, trials before commercial units, and scale gated on KPIs per use case. It is the template for buying a technology that is still improving.
Finance: First commercial deployment is 13 months away and the ROI timeline GXO gave in August has not moved. Apptronik, which raised $935 million this year, carries the platform risk; GXO buys an option.
ML engineers: Hands, speed and the brain remain the stated hurdles, the same bottlenecks J.P. Morgan named last week. A demanding operator paying per useful robot-hour is the contract that will produce the data on whether they are clearing.
Beginners: A giant warehouse operator will rent up to a thousand human-shaped robots from a US startup, but only as fast as the robots prove they can pick and pack to agreed standards, starting in 2027.

Source: Forbes · Full story

7 of 9

Fractile raises $720 million at a reported $7 billion, seven times its May valuation, before shipping its first inference chip; the FT says Anthropic wants to be an early customer

London-based Fractile closed a $720 million Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with Alphabet's CapitalG, DST, BlackRock, Bain Capital Ventures, T. Rowe Price, Shopify Ventures, Balderton, the UK's Sovereign AI fund and the British Business Bank participating. The Financial Times put the valuation at about $7 billion, against roughly $1 billion at a $220 million round in May. Fractile stacks standard DRAM directly on logic to attack memory bandwidth, which founder Walter Goodwin argues has grown 40-fold in 25 years while compute grew a millionfold; it claims up to ten times faster inference, has no commercial product, and plans its first chip for 2027. The FT reported Anthropic has expressed interest in being an early customer.

Finance: A 7x step-up in six months on a pre-silicon company is a bet that inference becomes a market as large as Nvidia's today, as Thrive's Philip Clark put it. The comparables are Cerebras, public since May, Nvidia's own inference chip from March, and Etched, MatX and Olix still private.
ML engineers: If memory-on-logic works, long-context and agentic workloads that are bandwidth-bound get the biggest gain. Nothing to buy until 2027; something to ask your inference provider about when they quote 2027 capacity.
Executives: The UK government is now a shareholder in a chip startup via its Sovereign AI fund. Expect procurement preference to follow investment, as it has with US champions.
Beginners: A British startup that has not yet built its chip raised $720 million from top investors, who value it at $7 billion, because it claims a design that moves data to the processor far faster than today's AI hardware.

Source: SiliconANGLE · The Next Web · Full story

8 of 9

ARTEX inside AWS: Datadog and Check Point publish the call sequence, 2,557 console hosts, ten user-agents and the finding that only Anthropic's model ever refused

Behind Friday's FBI alert sit two vendor reports with operational detail. Datadog Security Labs found that every ARTEX session it reviewed in October began with anonymous S3 reads of front-end assets, then split into two clusters: a stolen long-term key calling CloudWatch FilterLogEvents to hunt logs for crypto, hedge-fund and bank names plus 'Invoice', 'Billing' and 'AKIA', and self-registration on public Cognito user pools, some with the user-agent 'Mozilla/5.0 ARTeX', followed by a token exchange and an attempt at almost every read-only discovery call, all denied by a correctly scoped role. Check Point counted 2,557 hosts that had run ARTEX panels, 1,608 live, ten leaking their full databases; DeepSeek Flash models dominated, and the only guardrail trigger it saw came from Anthropic's Opus 5.5, after which the operator switched models. The framework speaks OpenAI and Anthropic API formats and swaps providers when a key is revoked.

Security: Hunt for the ten published user-agents and IPs, then for the sequence: anonymous S3 GetObject bursts, a long-term key calling FilterLogEvents across regions, Cognito SignUp from unfamiliar ranges followed by GetCredentialsForIdentity. Rotate long-lived keys; the one in Datadog's first cluster stayed valid for weeks.
Developers: Anything in a public bucket, including build manifests and runtime configs, is the agent's first read. Check every Cognito pool for open self-registration and read the authenticated role's policy.
ML engineers: Refusals bought minutes: one trigger, then a model swap. Provider-side safeguards are a speed bump in a multi-provider framework, which is why the useful controls are in the victim's cloud account.
Beginners: Security firms have published exactly what the AI hacking toolkit does once it is inside a cloud account, step by step. The attacks are old tricks run very fast; the defences are keys that expire and permissions that deny by default.

Source: Datadog Security Labs · Check Point Research summary via Let's Data Science · Full story

9 of 9

Corelight: 23 of 35 popular MCP servers resolve dependencies from public registries on every launch; one commit-pinned server ran 15 different dependency sets in five weeks

Corelight's James Lagermann measured how the MCP servers that AI assistants launch actually get their code and found that 23 of 35 use a recommended install, typically npx or uvx, that resolves dependencies from a public registry on every launch or build. A server pinned to one release and one commit ran with 15 distinct dependency sets over five weeks as 23 of its 69 transitive packages changed, including the cryptography library, a JWT library that moved through four versions and the MCP SDK itself; endpoint telemetry showed 17 to 21 rebuilds per server a month and one server launched 162 times in a week. Five frozen servers started 457 times matched their install-day package lists exactly. The Cloud Security Alliance's note pairs the finding with the March axios compromise and CVE-2025-6514 in mcp-remote.

Developers: Pin with package@x.y.z or --from 'tool==x.y.z', add a release-age cooldown (npm min-release-age, pnpm minimumReleaseAge, uv exclude-newer), set ignore-scripts, and prefer digest-referenced containers for anything with credentials. Treat MCP config files as code that needs review.
Security: A compromised transitive release reaches every developer and agent on the next launch, within hours. Route npx and uvx traffic through an internal artifact proxy so you have a log of what resolved and a place to block it.
Executives: Your coding assistants are re-installing their tools from the public internet many times a day. That is a supply-chain surface nobody budgeted for; one control, an internal registry proxy, covers most of it.
ML engineers: The approval a user gives an MCP server binds to a name, not to the content that runs. Assistant vendors should bind approval to a digest; until they do, pin on your side.
Beginners: The plug-ins that let AI assistants use tools quietly download fresh code every time they start, so a poisoned update anywhere upstream spreads fast. The fix is to lock the exact versions, as you would for any software.

Source: Cloud Security Alliance research note · Corelight blog · Full story

AI Briefing · summaries written in our own words and checked against the linked reporting · Subscribe · Archive · corrections: corrections@aibriefing.io