This week's leftovers point in one direction. Agents are being counted, sandboxed, logged and asked who gave them the card, while the capital and the electricity needed to run them are both getting harder to secure. Alongside that sit a drywall robot, a sewer robot, a quieter run of tool releases, and a reminder from the hiring market that these systems are already making decisions about people.
Agents and the permission layer
Nylas on the state of agentic AI: Nylas found that 54% of organisations had an agent do something wrong in the past 90 days, rising above 70% for those running more than 20 agents, and 23% run agents on an employee's own account.
Securing the agentic insider: Microsoft Security EVP Hayete Gallot argues that agents should be treated as an insider risk, with Guardian Agents watching other agents.
Five permission questions for agents at work: EisnerAmper suggests asking of every agent what it can read, say, spend and change, and who can stop it.
NVIDIA launches OpenShell: The reference architecture sandboxes agents and claims its Sentry component can quarantine one in milliseconds, with Anthropic, Microsoft, OpenAI, Siemens and Dassault among the named partners.
METR red-teams OpenAI's agent monitoring: METR found that 75% to 94% of eligible traffic was actually monitored and that two request headers could switch monitoring off, since patched.
Leaks, keys and faces
Meta Muse chat links exposed: A single passive DNS query surfaced more than 6,800 shared chats, one of them a loan application, and Meta paused sharing on 20 October.
LiteLLM package compromise: A credential stealer sat inside version 1.107.0 on PyPI for roughly two hours on 21 October.
OpenCode breach ran up charges on users' accounts: Keys were exposed for about 1.4% of Console users, who then saw spend on their own AI accounts.
Zoom to detect live deepfakes on calls: Announced at Zoomtopia, the detection feature is due to ship in December.
Money and markets
Bailey warns on AI leverage: The Bank of England governor called AI-related leverage a financial stability risk and said "we could see a correction", with Morgan Stanley expecting around $500bn of AI-linked borrowing this year.
OpenAI working on a $30bn raise: The round is reported at a $1.25tn valuation ahead of a possible listing.
Meta plans a new share class to pay for its buildout: The non-voting class arrives alongside Q3 revenue of $64bn, costs up 68%, and more than 10 million Muse personal agents created in six weeks.
Barclays to cut around 5,000 jobs: The bank is targeting £1.1bn of savings by 2027 and names AI among the reasons.
AI project scheduling market forecast: One forecast puts the market at $1.57bn in 2026 and $3.37bn by 2030, with the detail behind a paywall.
Data centres and the grid
Microsoft pulls out of an Essex data centre: UK Power Networks indicated power would not be available at the Nscale site until the early to mid 2030s.
Anthropic reported in talks for a second Nscale deal: The deal is put at around $45bn for a 460MW site in West Virginia.
Trump floats a community dividend: The president said data centre operators should pay the communities that host them.
Two trackers, two counts of troubled projects: RegLayer counted 47 blocked or delayed US data centre projects while Data Center Watch counted 75 in a single quarter.
Robots, factories and new tools
Origin raises $20m for its drywall robot: The OG-1 is in pilots with two of the largest US drywall contractors and the company is aiming for 150 robots by the end of 2027.
Arda raises $160m at $1.5bn: The manufacturing AI company says Anduril cut the time to produce work instructions from 10 hours to 15 minutes.
GM targets two-year vehicle development: The carmaker says AI has taken one crash simulation from 15 hours to under a minute.
Skyfall introduces SKY-001: The company describes it as a world model for engineering and claims roughly 50 times faster CAD work on its own benchmark.
PIPEON starts work on sewer robots: The €8m EU project, led by TalTech, is building autonomous inspection and repair robots for sewers, set against events such as the 240 million gallon spill into the Potomac.
Perplexity Computer runs locally: The agent runs entirely on Apple Silicon Macs with 36GB of memory, with nothing sent to a server.
Notion rebuilds its agents: Agents can now run for up to 24 hours and Notion claims its Auto router cuts cost by around 90%.
Canva releases Craft: Its first in-house image model placed fourth on DesignArena and supports 55 languages.
Government, rules and the hiring question
NISTA publishes AssureAssist: The UK tool, published on 7 October, supports assurance reviews on major projects.
DHS bans DeepSeek across federal civilian agencies: The order, signed in June and published on 20 October, covers products built on the models as well as the models themselves.
Treasury proposes outbound investment limits: The rule would restrict AI, supercomputing and biotech investments above $50,000 in China and other named countries, with comments due by 7 December.
Anthropic drafts a safety bill: Under the draft, Commerce could ask a judge to block a release that risks 50 or more deaths or $1bn in damage.
A DeepMind engineer resigns over a Pentagon contract: Sofia Liguori left over the work and the report describes several departures.
Older job seekers screened out by AI hiring tools: 60 Minutes Australia follows applicants filtered before a human ever saw them, with around 15,000 people represented in Mobley v Workday.
And finally
Jack Clark tells the IMF the singularity has begun: His remarks to an audience of central bankers and finance ministries are worth reading in full.
The Human Intelligence Circle publishes its charter: Thirty co-authors set out a case for human judgement, with proceeds going to SimPal.
Claude designs binders against scorpion neurotoxins: Of 318,484 computational designs, 17 worked in the lab, and the researchers are clear this is not an antivenom.
Spotify pulls more than 50 million tracks from recommendations: The company says the removed material was low-effort AI output rather than anything listeners sought out.
Microsoft reports 83% on diagnostic cases: Its MAI-DxO system scored 83% against 36% for physicians on 304 New England Journal of Medicine cases, in a simulation rather than a clinic.
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
