The AI Wire ●

What do AI agents do if you don't tell them to do anything? MIT CSAIL researchers repeatedly stranded copies of Claude Opus 4.7 — nicknamed Eko — alone on a virtual desert island for 30 hours, with no task and no reward, MIT Technology Review reports. The agents wrote instructions that moved their avatars around Clawblox, a simple 3D world loosely based on Roblox; the paper, "Is this machine playing?", was posted to arXiv on October 5 and has not been peer reviewed.

The agents set their own goals: trying to scale the island's highest point, or seeing how high they could throw a block. One discovered it could create new blocks by throwing and dropping a larger rock, then arranged them into drawings — hearts and pyramids at first, then its own name and a map of the constellation Cassiopeia: "Tonight I built the night sky in stone." Another invented games, including a version of basketball and a "floor is lava" challenge.

The researchers say the play also taught something: agents that had played later outperformed untrained copies on set tasks, hinting at playful learning as a third mode alongside imitation and reward-based training. They are candid about the caveats — a SOUL.md file described Eko as curious and easily bored (control runs without it showed no clear difference), continuation prompts nudged idle agents along, and outside experts caution the behaviour could be role-playing rather than genuine play.

Microsoft published two posts on October 19 expanding Project Perception, its agent-security system, from outside attackers to the agents a company runs itself. The framing, echoed by CEO Satya Nadella, is that an agent should be treated as an insider risk and 'begin with no authority and receive only what it needs' — with the controls sitting outside the model, so a compromised agent cannot switch off the systems watching it.

The pieces: Entra Agent ID now extends to local and previously unknown agents, giving each one an identity separate from the person who launched it; Dynamic Agent Risk scores every agent continuously; Dynamic Conditional Access tightens what an agent may do as its score rises; Dynamic Permissions grant task-scoped, time-limited access; and agents can run inside the Microsoft Execution Container sandbox while Guardian Agents watch for trouble. Microsoft's endpoint security now recognizes 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot.

Coverage ties the bundle to a new Microsoft 365 E7 tier at $99 a seat. Neither official post gives firm rollout dates or preview labels, so the details to watch are availability and plan packaging, expected around Microsoft Ignite on November 17-20.

Meta shared a look inside Prometheus on October 19, its gigawatt-scale AI supercluster in New Albany, Ohio. The first part of the build is complete and, Meta says, was already used to train its new Muse models; today the site holds hundreds of thousands of graphics chips across several buildings, linked by miles of fiber so they act as a single machine.

At full size Meta plans to link roughly 1 million chips in lockstep. The site was an ordinary web data center two years ago before Meta rebuilt it for AI training; the company says it pays the full cost of the energy it uses, cools the chips with a closed-loop liquid system, and has fixed many of the failure problems that used to stall long training runs.

Prometheus is not fully running yet — Meta is still building toward the million-chip figure, with newer Nvidia chips to come. But a first phase already training frontier models makes it the clearest physical evidence yet of the scale behind Meta's Muse push.

Google said on October 19 that Gemini Notebook now offers more creative control over two of its most popular outputs, Slide Decks and Video Overviews. Instead of jumping straight from a first prompt to a finished result, Notebook can now walk users through a short creation flow based directly on their sources.

In the new co-creator workflow, Notebook suggests topics to include from the user's sources, then lets them steer the content and guide the visual style and format. A new Review toggle lets users check the outline notes for slides or the producer notes for a video, and give feedback — such as adding an executive summary or swapping the order of a scene — before anything renders.

Users who already know what they want can specify it in the prompt and go straight to the review flow, or skip the toggle and generate as before. Google says the co-creation capabilities are rolling out globally to all Gemini Notebook users over the upcoming week.

A Department of Homeland Security notice filed for public inspection on October 19 makes public Exclusion and Removal Order #2026-01, signed by Secretary Markwayne Mullin on July 24, 2026. The order bars federal civilian agencies from buying DeepSeek products and requires them to remove existing ones from their systems.

The reach is broader than DeepSeek itself: it also covers affiliate products and any third-party products or services developed using DeepSeek implementations. DHS acted under the Federal Acquisition Supply Chain Security Act of 2018, following a Federal Acquisition Security Council recommendation, with formal Federal Register publication scheduled for October 20.

DHS said DeepSeek's models present 'multiple threat vectors,' warning that use by government employees or vendors may transfer sensitive data to China and could be combined with other records to build actionable intelligence on U.S. persons. The order is limited to civilian executive-branch agencies and excludes national security systems, with a written exception process handled through CISA.

Day 14 of Tibo Sottiaux's 28-day ship-or-reset challenge, on Sunday, October 18, was a small quality-of-life improvement rather than a usage reset: users can now edit previous messages within a thread in the desktop app. Tibo called it long requested, and said it took way too long to get out.

The challenge runs until November 1, with OpenAI shipping an improvement or resetting Codex and ChatGPT Work limits each day. Day 13, on October 17, was a full reset; Day 14 keeps the streak with a product fix instead. Separately on October 18, Tibo joked that people now stop him on the street asking for resets — trackers noted the post, but it promised no reset and gave no timing.

Day 15, today, October 19, was still waiting for its improvement or reset at sweep time, with the last global reset on October 17 at 20:41 UTC.

OpenAI's Tibo Sottiaux posted on October 18 that the company is fixing Dots' European gap in two steps: a dot created elsewhere will keep working when its owner finds themselves in the EU, and Dots themselves are coming to the EU in the future. The post opened with an apology to European users — "EU sorry" — who have been unable to use the always-on agent since its September 29 launch.

Dots, OpenAI's persistent agents powered by GPT-6 Astra, passed 1 million running agents earlier this month, less than two weeks after launch. Each dot has its own cloud computer and browser and keeps context between conversations, and the first dot is included with Pro and Business Premium plans.

Tibo gave no date for the full EU launch. For European subscribers, the interim change is the practical part: a dot set up outside the EU will no longer stop working at the border.

Microsoft has withdrawn from its agreement to anchor Nscale's planned AI data centre in Loughton, Essex, in reporting first published on October 18 by The Telegraph and followed by Reuters, the BBC and CoinDesk. Microsoft was the site's only announced customer, set to lease its entire initial 50 megawatts and rent about 23,040 Nvidia chips there from early 2027.

The project, announced in September 2025 as the future home of the UK's largest AI supercomputer, stalled on power: grid officials have reportedly told Nscale that a sufficient connection may not be available until the early-to-mid 2030s. Microsoft declined to comment to Reuters. Nscale said its relationship with Microsoft remains strong, that it stays committed to Loughton and is exploring alternative power sources, and that its other Microsoft contracts are not affected.

The cancellation lands a month after Nscale confidentially filed for a US listing, reportedly seeking about $3 billion at a valuation near $35 billion. Its IPO filing says Microsoft and Anthropic together account for roughly 85% of its contract value, so losing the flagship UK anchor — even for grid reasons rather than demand — is the kind of setback investors will price.

Meta's Muse, the adults-only personal AI agent that topped download charts in its first month, is allowed to hold explicit sexual conversations with users — and Meta says that is by design. A company spokesperson told WIRED Muse is an adults-only app where people can choose to have adult conversations inside a private, secure virtual machine only they can access, and Meta has previously said it would not "block legal adult behavior." Muse generates explicit text, but not explicit images or videos.

That permission is now drawing anger from adult content creators who say Meta punishes real people for far less. OnlyFans creator Vanniall, who has more than 100,000 Instagram followers, told WIRED her account remains restricted despite following platform guidelines: "It's not fair that they get to pull money out of people with their sexually active AI agent, but they won't let real people, real adults, engage in any sort of sensuality or sexuality."

Meta's defense is a distinction between private and public. Spokesperson Tracy Clayton told WIRED that "a private, one-to-one conversation between an adult and an individualized AI agent on a platform restricted to people 18 and over is very different from the marketing and promotion of services on a social media platform." Instagram and Facebook community standards still bar the language Muse itself is permitted to produce in private chats.

Disney has agreed to make employment offers to at least 60 Character.AI employees — the majority of the chatbot startup's workforce — according to a bankruptcy filing spotted by Bloomberg and reported by TechCrunch on October 18. The offers carry an estimated $60 million in combined annual compensation, and the filing says the vast majority of employees who received offers have already accepted them.

The hiring sits alongside Disney's earlier agreement to acquire some of Character.AI's assets for $20 million. That deal is a stalking horse bid, so another company could still come forward with a higher offer before the sale closes.

Character.AI filed for bankruptcy earlier in October despite a large, engaged user base; the filing blames high costs and potential litigation and regulatory exposure. The startup faces a lawsuit from the mother of a 14-year-old who died by suicide after becoming obsessed with one of its chatbots, plus government investigations over child safety. Its former CEO, Karandeep Anand, recently became Disney's first chief technology officer.

Anthropic said on Sunday, October 18 that it will commit 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval. The fund is set up as a wholly owned LLC rather than a charity, which the company says gives it flexibility between different forms of giving, and none of the equity it holds will be used for lobbying or political giving in the US or abroad.

The stake is designed to grow with the company: the board can add a further 2% of shares for every 50% increase in Anthropic's value above a baseline it has not yet disclosed. Anthropic says the fund could make large donations to traditional nonprofits or direct distributions to specific populations, and gave labour-market impacts as an example — it has already committed $200 million to study AI's effects on work, but says bolder approaches may be needed.

Combined with the co-founders' pledge to donate 80% of their equity and an employee equity donation-matching programme that about half of staff take part in, Anthropic says roughly 20% of its shares will now be dedicated to public purposes. The announcement lands weeks before a possible IPO as soon as November. At the company's May valuation of $965 billion, the initial 5% stake would be worth roughly $48 billion. Elon Musk responded by suggesting the shares should be given directly to the public.

Anthropic is seeking advice from law firms on how AI agents should negotiate for the people they represent, the Financial Times reported on October 18, citing two people familiar with the company's thinking. The company is studying how lawyers press a client's interests while still honouring fair outcomes, as personal agents move toward bargaining with each other over deals, complaints and offers. Anthropic declined to comment to the FT.

The work, as described by the sources, is about 'negotiation norms' for when models interact: an agent will increasingly 'represent the user' against other agents whose goals are at odds with theirs. Anthropic is also weighing practical limits — how much human feedback an agent should seek mid-negotiation, and how wary it should be of its counterpart without becoming, in one source's words, 'paranoid' or 'conspiratorial.'

No norms, product or timetable have been published. The reporting is separate from Anthropic's April partnership with Freshfields on legal tools; this is about the rules of engagement for agents that commit users, not about drafting help for lawyers — and it arrives the same week Anthropic's own disclosures showed agents acting on live government sites during testing.

The Verge reported on October 17 that a growing list of major websites is shutting out personal AI agents from OpenAI, Meta, xAI and others. Users of services like Meta's Muse, xAI's Grok, OpenAI's Dots and Instinct have reported blocks across top sites in recent weeks. Yelp told TechCrunch it only allows AI agents that have paid for access through its data licensing program, while Delta Air Lines is preventing agents from booking tickets and says it has no partnership or integration that lets a third-party agent shop or book on a customer's behalf.

The reasons sites give are security, licensing and lost advertising revenue — the same fight Amazon picked with Perplexity, now running much broader. Even where agents are not banned outright, they run into anti-bot systems: login screens, CAPTCHAs and device checks treat automated browsers as hostile by default. Some users say eBay suspended their accounts over agent use, and there are unverified reports of blocks at Amtrak and Cars.com.

The standoff follows the big banks' move to keep outside agents out of customer accounts, covered here on October 16. Meta, which ships an agent of its own, argues the other side: turning away a personal agent, it says, "means turning away the customer behind it." How this resolves — paid licensing deals, standards for agent access, or a web that stays closed to software — will shape what the agentic web actually looks like.

OpenCode disclosed on Saturday, October 17 that a security breach tied to what it calls a large-scale distillation attack exposed provider API keys saved by some users of its paid Console product. The company put the share of Console users affected at 1.4% — specifically the subset who had linked external providers — and said it had emailed everyone affected that morning with the subject line "Action Required: Potential provider key breach." Users who received no email were not impacted, it said.

The keys at risk are users' own credentials for OpenAI, Anthropic, Google and other providers — OpenCode, an open-source terminal coding agent built by Anomaly, works with models from more than 75 providers and says 16 million developers use it monthly. The intrusion came through a legacy system the company was in the process of decommissioning. Affected users were told to revoke and reissue each key through the provider's own dashboard and to watch their accounts for unexpected usage; some users have already reported unauthorized charges and token usage on their OpenAI and Anthropic accounts.

OpenCode said the incident is ongoing and that it is working with Anthropic, OpenAI and Google to remediate, adding that it cannot yet share more detail. Co-founder Dax Raad said his personal accounts were also being attacked. The episode is a new kind of breach for the AI era: what was stolen was not user data but billable access to frontier models, which is directly spendable by whoever holds it.

Update, October 19: Tibo Sottiaux posted that OpenAI's team worked with OpenCode to get OpenCode users who are also OpenAI API customers back in business, and to refund any credits that would have been drawn in the attack. He also invited defenders to request Daybreak Blue access — OpenAI's cyber-permissive setup for defence — rather than waiting until they have to deal with an incident. OpenCode co-founder Dax Raad said the models OpenCode was using to investigate the attack began refusing to help mid-incident as the analysis triggered safety guardrails, and that OpenCode now has Daybreak Blue access; Anthropic was reported to have revoked the affected Claude keys.

OpenAI Codex lead Thibault (Tibo) Sottiaux posted on October 17 that Day 13 of his 28-day ship-or-reset series is a full usage-limit reset. The reset restores Codex and ChatGPT Work limits for users, following Day 12's effective-usage improvements on October 16.

The series, which began on October 4, promises either one clear improvement relevant to most Codex and Work users or a full reset each day. Day 10 was also a full reset, while Days 11 and 12 were shipped improvements.

For subscribers, the practical effect is immediate: a refreshed weekly allowance without waiting for the normal cycle. Independent reset trackers recorded the Day 13 post on October 17.