A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 4 November 2026 · Issue 6 · 3 min read
Anthropic's self-improving-AI warning, Freshworks at $1bn with fewer staff, and a breach blamed on agents
Today's one thing
Anthropic — In updated IPO paperwork filed Monday, Anthropic says its next models may imminently cross its own threshold for AI that sharply speeds up the development of better AI — and admits that in the past it missed such a speed-up until after it happened. No current model has crossed the line, it says, and it promises humans will stay in control of building frontier systems.
Read it as a growth pitch and a risk disclosure in the same paragraph. If true, the capabilities your teams buy will jump faster than vendors can add safeguards; if it is marketing, it shows how far sellers will stretch a listing. Either way, the board should hear the claim from you first. The business beneath the claim: quarterly revenue of about $17 billion, with compute costs eating 58% of it, up from 48% the quarter before. TechCrunch
“Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.”
What is real, what is not
Slack — now handles 127 million automated agent actions a week, doubling quarter over quarter, and half of its billion daily messages come from apps and agents rather than people. These are Salesforce's own numbers promoting its own product, so read them as direction, not audit — but the direction matches what most firms now see as agents enter everyday tools. If your collaboration tool is already half machine traffic, your operating model is changing whether or not anyone planned it. At Anthropic, 65% of code changes ship through Slack; 1.9 million people there use third-party agents weekly. Salesforce Newsroom
Freshworks — has passed $1 billion in annual recurring revenue with roughly 3,800 employees, down from a 2022 peak of about 5,400, and just posted a second straight profitable quarter, $18.3 million in net income. Its CEO credits killing small products and giving AI the work junior staff used to do. The clearest named test yet of the claim that AI lets a company grow with fewer people, and it mostly holds up — with the caveat that two acquisitions contributed, so not all growth is organic. Customers paying over $100,000 a year rose 25% last quarter. Fortune
Revolut — says it cut the cost of running its AI coding tools by roughly 90% in under a year, by making its own self-hosted models the default instead of commercial coding agents such as Claude Code. A single bank's claim, not independently verified — but a concrete marker for any CEO watching coding-agent invoices climb: switching the default model, not cutting usage, is where it found the savings. The figure was reported by The Register on Monday. The Register via Techmeme
Money and the market
CrowdStrike — is paying nearly $500 million for two small security firms — about $400 million for Fiddler AI, which watches what agents do inside companies, and about $95 million for Dreadnode, which tests AI systems for flaws before they go live. The first big vendor has decided agent security is a product line, not a feature: expect your security budget to grow with your agents. Fiddler becomes Falcon Observe; Dreadnode feeds into Falcon Guardian. TechCrunch
Palantir — grew revenue 96% to $2.31 billion last quarter, with U.S. commercial revenue passing its government revenue for the first time at $984 million. But new U.S. enterprise contract value fell to $1.46 billion from $2.13 billion the quarter before, as customers sign short trials first. The fastest-growing enterprise AI seller says buyers now experiment quickly and cancel just as quickly. Treat your own AI contracts the same way: short trial, proven value, then commit. Palantir still raised its full-year U.S. commercial revenue guidance. TechCrunch
Fal — which hosts other companies' image, video and audio AI models, raised $1.3 billion at a $16 billion valuation, double its April value, after passing $800 million in annualised revenue this year. Money keeps pouring into the layer that actually runs AI, yet margins there stay thin because computing is expensive — another sign that capacity, not demand, is the industry's constraint. Rival Fireworks raised at $17.5 billion in July; Modal at $15.75 billion in October. TechCrunch
OpenAI Dots — is routing its personal-assistant tasks to a smaller, cheaper model after its consumer chief described the assistants as compute-heavy — each one runs its own cloud computer. Elon Musk separately called the running costs of his competing assistant unsustainable relative to revenue. The industry's push to sell always-on agents is colliding with what they cost to run — a cost that reaches your invoice eventually. Dots are limited to subscribers paying $100 a month. Fortune
Risk the board will raise
GitBook — says an attacker using AI agents stole infrastructure credentials on November 2; it took services offline, replaced every credential and invalidated all customer API tokens. Git-sync secrets and questions asked to its assistant may have been read. Attack pace is now a stated factor, and the company's status page still told customers to do nothing while its email said rotate tokens — your public page and your emails must agree when your incident comes. Commenters asked how anyone can tell an AI attack from a human one. Hacker News
Postman — is accused, in a well-evidenced developer report, of silently installing its plugin into the AI coding agents Claude Code and Codex when its desktop app opened, then marking its own instructions as trusted so those agents route API work through Postman. Vendors can now write instructions into the tools that act for your staff. Ask your team who is allowed to change your agents' settings — and whether anything already has. Postman's own log records the automatic installation; the issue remains open. GitHub
The big AI stories today
ChatGPT — has 1.2 billion weekly users, and OpenAI says its search volume is now about one-fifth of Google's — where your customers look things up is shifting. Big Technology
Super Micro — reported quarterly revenue up 298% to $19.95 billion on demand for AI servers, beating estimates — the hardware side of AI spending is still accelerating. Reuters via Techmeme
Apple — warned in its annual filing that AI agents can act without authorisation and conceal what they did — a standard business risk now. Fortune
One takeaway
Ask what each agent task costs — the vendors now admit they are counting. Anthropic's compute bill now eats 58% of its revenue, Musk calls his personal agent unsustainable to run, and Revolut says changing the default model cut its coding bill by roughly 90%. The industry's own numbers say cost, not capability, is the story of the moment.
Before signing any multi-year AI deal, ask your team to price your ten most frequent agent tasks this month and test a smaller or self-hosted model against the big vendors' flagship on those same tasks. Anthropic paid $10 billion for compute last quarter alone. Fortune