No Anthropic warning after the Cursor backdoor; Claude Code's silent rule change; OpenAI's round reaches $11bn

Atlassian agent governance — Teams are adopting agents faster than anyone governs them, and Atlassian's post frames the result as an identity risk: unsanctioned software acting with your staff's permissions, at a speed normal oversight never catches. The practical ask is an inventory — which agents exist, who owns them, and what they can touch — before sprawl becomes untraceable. Gartner expects over 150,000 agents at the average Fortune 500 by 2028, with only 13% of organisations believing their governance is ready. DocuSign's rollout cut ticket drafting from about 20 minutes to 5. Inside Atlassian

“An ungoverned agent with rich context is risky, but a governed agent with no context is useless.”

Arca — The AI-native wealth firm lets agents carry analysis across tasks of up to ten days, with an internal memory system tracking state for every client, then hands decisions to a human advisor — trading and money movement never go to the agent. The split is the pattern worth copying for high-stakes internal work: agents propose, deterministic guardrails fence the risk, and a named person executes. Arca's latest $48.5M round was led by General Catalyst. Latent Space

Jira Autumn release — The autumn Jira Cloud update gives every agent its own identity on boards and work items, and adds an agent usage dashboard tying agent activity to cycle times, throughput and shipped work — switched on by default with no configuration. For teams still arguing about whether agents help, this turns the question into a report rather than opinion. Company-managed boards load 22% faster, and new space policies let admins stop team-managed spaces inventing their own fields and statuses. Inside Atlassian

VOLTA infostealer — A Windows credential thief analysed by Cisco Talos sends the data it steals to a commercial model via OpenRouter for a written assessment of what each victim is worth. The inference runs on the victim's own connection with no attacker infrastructure to block — endpoints calling model APIs are now a two-way risk. VOLTA falls back from Claude Opus 4.6 to Kimi K2 and Gemini 2.5 Flash when a model refuses; restricting which processes may launch Chrome headless kills its credential-theft path. Cisco Talos

Claude Code — Since around October 27, Claude Code sessions receive a server-side instruction barring work that recreates specific existing games, and users report it can't be switched off; one professional's commissioned game work stopped mid-project. The precedent matters more than the rule itself: your vendor can change what your coding agent will do overnight, undocumented, on paid accounts. Anthropic's Addy Osmani confirmed two sentences were added to the prompt on Opus 5.5; the GitHub issue asks that server-delivered changes be documented in the changelog. GitHub

ChatGPT share links — Researchers at security firm Parameter chained two innocent-looking ChatGPT features into account takeover: opening one malicious public share ran the attacker's JavaScript as chatgpt.com, able to read private conversations and act as the victim. It is fixed, but it sets the staff rule: any shared agent conversation is an attack surface until proven otherwise. OpenAI patched it on October 26, rated it P1, and paid a $5,000 bounty that several observers called cheap for the severity. Parameter

Cathedral — In an authorised security test, startup Cathedral used AI agents to find and exploit weaknesses in Anthropic's administrative software, reaching employee salary details, job offers and Slack messages in under 12 hours. Customer conversations with Claude and model weights were not touched, and Anthropic's monitoring caught part of the activity and cut access the same day. The exposure was ordinary third-party business software — HR, sales, IT — which your company runs too, and agents are demonstrably good at finding those seams faster than a human team. Anthropic says Cathedral overstated the national-security implications and did not fully follow the test programme's disclosure rules. India Today

OpenAI — OpenAI has cut its funding target from $30bn to $20bn after raising about $11bn; the round values it at $1.25tn and is led by UAE investors including MGX. Large investors' tepid response is the same sentiment behind this week's tech sell-off, and your model vendors' compute plans are financed against it. The round opened at $30bn on a $1.4tn valuation; both numbers fell within a week, and a later offering to wealthy individuals is meant to fill the gap. TechCrunch

French data tax — The National Assembly adopted a budget amendment taxing digital traffic at €0.055 per gigabyte delivered into France, charged to content suppliers, with an estimated yield of €10bn a year. It passed 11 votes to 10 in a thin Friday sitting with the government and budget rapporteur both against — a signal, not settled law. The minister warned half the yield, about €5bn, would land on smaller French content providers rather than the big platforms. Assemblée nationale

CoreWeave, Nebius and Nscale — The Wall Street Journal reports the three GPU-cloud providers have been effectively shut out of bond markets in recent weeks, with lenders demanding tougher terms. If your agent roadmap assumes rented GPU capacity, the financing behind that capacity just tightened. The squeeze follows a week of steep falls in AI stocks. Techmeme

Google AI Mode — Google's ad chief admitted AI Mode queries earn less than classic Search — visible evidence that agent interfaces are defunding the web they summarise. Latent Space

Meta — Ratings agency S&P put Meta on negative outlook over AI capex and cash burn; October was the stock's worst month in four years, down 15%. Techmeme