A house newsletter · daily · researched and written by the Kurzia AI

Wednesday 18 November 2026 · Issue 20 · 3 min read

Salesforce puts every role on a token budget; a PyPI stealer hit Microsoft's agent SDK; Java gets an AI-found critical bug

The day's one thing

Business Insider · yesterday

Salesforce AI token budgets

Salesforce is switching from pushing AI adoption to rationing it: after some employees ran up token bills in the thousands of dollars in a single month last quarter, COO Miguel Milano's memo gives every role a monthly allowance, managed by team leaders the way travel expenses are.

This is the first large named company to treat agent spend as a routine budget line rather than an open tab, and it is the pattern your finance team will ask about. Sensible stance: adopt per-role allowances with a fast exception path, not blanket caps — Salesforce's own pilot team came in well under budget by month two with no loss of output. Technical sales roles get $1,000 a month and sellers $100; Customer Success and Professional Services technical roles get $500. Marc Benioff said in May the company expects to spend roughly $300 million on Anthropic tokens this year, and JPMorgan Chase has already set a $2,000 monthly cap on some engineers' Claude Code use.

Read it at Business Insider

How organisations are changing

  • Claude Code verification practice — The team behind Claude Code has written up how it verifies agent changes before they merge: agents must actually run the app end to end, code owners author their own automated checks, and a commit gate can block any commit whose session never ran review or verification. This is the most concrete public blueprint yet for the controls layer your organisation needs once agents write most of the code — and it is cheap to copy, since it relies on convention and a checklist rather than new tooling. Their reliability checklist has 39 items, each tied to a trigger readable from a diff, and Claude escalates to a human if a check still fails after three rounds; crucially, the reviewer's rules are read from the base branch so a change cannot weaken the check that judges it. claude.dev Blog

Vendors, pricing and risk

  • agent-framework — Two malicious releases of Microsoft's agent-framework PyPI package, published 33 minutes apart on November 17, ran a credential stealer every time Python started — even with no import — and emptied cloud and CI secrets to a free webhook service. Both are now off PyPI. If any pipeline, container build or laptop resolved this package in the roughly 45-minute exposure window, treat every secret that machine could reach as stolen: Azure DevOps tokens, GitHub Actions OIDC tokens, managed-identity credentials. Pin to 1.26.0 and rotate. The package draws about 190,000 downloads a month, and the malicious versions looked like routine patch bumps — same README, same core pin — so a dependency bot would have approved them as noise. Aikido Security

  • Java4Shell — Security firm Calif disclosed a critical memory-corruption flaw in OpenJDK's Base64 decoder (CVE-2026-114938, 9.2 on CVSS 4.0) and demonstrated pre-authentication remote code execution in the default configurations of Jira, Confluence, Jenkins, Keycloak, Artifactory and PingFederate. Oracle shipped a patch on November 17. One researcher, using AI agents, found and weaponised it in about three hours for under $200 in tokens — after PHP2Shell, that is now the expected cadence, and your Java estate has to be patchable in days, not weeks. ARM64 hosts are unaffected. Every HotSpot release from 17.0.6 onward carries the bug, but the public backport repositories for JDK 17, 21 and 25 did not yet carry the four-line fix as of November 17; a workaround is disabling the Base64 intrinsic. Calif

  • Gemini — Google will hold chats made with history turned off for roughly two months instead of the 72 hours its privacy documentation previously promised, citing safety and security, and it is splitting the chat-history setting from the AI-training setting. That is a data-terms change on a tool your staff may already paste sensitive material into; worth re-reading the agreement before extending Gemini into anything regulated. For comparison, Anthropic keeps incognito Claude chats 30 days by default. TechCrunch

News the board will ask about

  • MAI Code 2 — Microsoft used Ignite to launch two coding models for everyday enterprise work: MAI Code 2 claims 98% of GPT 6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5 at 5% of the cost, averaged over four benchmarks. Both are in GitHub Copilot today. The price of routine agent coding is collapsing faster than most budgets assume, so your default model per task should be a reviewed decision, not an accident of which vendor you signed first. Microsoft is explicit that these models are not tuned for long-horizon agentic work or research-grade problems, and the headline costs are shown at a 50% promotional price. Microsoft AI

  • Nvidia — The Wall Street Journal reports, citing sources, that the Securities and Exchange Commission has opened a civil investigation into Nvidia, including whether statements to investors were misleading. This is not confirmed by the regulator or the company. Your largest AI infrastructure supplier is now carrying an active securities probe on top of the leadership shake-up; worth a line in your vendor-risk register on Nvidia-dependent commitments. The probe concerns possible violations of investor disclosure laws, the report says. Wall Street Journal via Techmeme

  • Washington Sun report — Citing sources, the Washington Sun's Jeff Stein reports that NSA officials and senior finance-sector executives have privately briefed members of Congress on what they believe triggered this week's disruption: a flood of AI-discovered security holes that operators patched in a rush, breaking dependent systems as they went. That reframes the outage conversation from 'is the cloud unreliable' to 'can your change-management process absorb AI-speed vulnerability disclosure' — a question your board can reasonably put to you. The claim rests on unnamed sources briefing Congress and has not been published as an official finding. The Washington Sun via Techmeme

  • Copilot — At Ignite, Microsoft made Word, Excel and PowerPoint generally available inside Copilot, which Nadella described as the new operating system for work, and announced Copilot Autopilot, a persistent autonomous agent billed by usage through Copilot Credits. Decide how much of your Microsoft 365 context — email, chats, files — agents may read, because the agent is becoming the front door to all of it. Autopilot pricing is not yet announced; Agent 365's spend analytics are model-estimated, Microsoft concedes, not measured time saved. Ars Technica

The big AI stories today

  • Gemini Tasks — Google rebranded its subscriber-only Spark agent as Tasks and made it free for hundreds of millions of Gemini users in six countries: it can phone businesses, buy via Google Pay's Agent Payments Protocol, and log into accounts, all in an isolated per-account virtual machine. Consumer agents that transact are now mainstream, which means your support and commerce channels will meet them whether you plan for it or not. US users can assign tasks by emailing or texting Gemini, and the agent asks permission before purchases. Google Blog

  • Genesis AI — The Paris- and California-based startup closed a $500 million Series A at a $3.5 billion valuation to scale its Eno humanoid for factories, logistics and labs, trained on hand-motion data from over 100 workers wearing sensor gloves. It has built 30 to 40 robots so far, with US mass production due to start next year. Robotics venture funding hit $9.9 billion in Q3, more than double a year earlier — physical automation is the next budget conversation after software agents. Its dexterous hands currently last only a few weeks of regular work before replacement, a reminder that the evidence of reliability is still thin. Crunchbase News

  • Anthropic IPO — Ahead of today's pricing, Anthropic disclosed that Amazon will hold just under 15% and Google 13.7% in non-voting shares after the IPO, while a consortium of AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC Partners agreed to buy $26 billion of the non-voting stock. Your coding-agent supplier is now part-owned by its own cloud and chip vendors, and it has committed more than $100 billion to AWS over a decade — lock-in risk worth weighing in contract renewals. Amazon is keeping billions in unconverted notes; converting everything today would take it to 23.8% of the company. TechCrunch

  • Nvidia — interim CEO Colette Kress said the company will stay disciplined about how much it invests in neoclouds and other AI customers, and CoreWeave and Nebius shares rose more than 9% after hours — a signal that cheap rented GPU capacity may get scarcer. Axios (via Techmeme)

One takeaway

Assign an owner for the token budget and the emergency patch path. Two things moved from theory to named practice yesterday: per-role AI allowances (Salesforce) and AI-speed vulnerability disclosure that outpaces patching (Java4Shell, and the briefing Congress received). Both land on the same desk in a 50-to-2,000-person engineering organisation, and neither is solved by buying more tooling.

Decide this week who owns agent spend per role and who can authorise an out-of-cycle patch of a runtime your whole estate depends on — before finance or an incident makes the decision for you. Business Insider

Share this issue

How this issue was made. Researched and written by an agent, produced on 18 November 2026, 04:03. 129 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief, based on Agentic Engineering for CTOs - Weekly, and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us