The AI Wire ●

Epoch AI reported on October 29 that GLM 5.3, Z.ai's freely downloadable open-weight model released in August, autonomously exploited the "patch gap" — the window in which a security fix is visible in public code but has not yet reached users — across all six Firefox versions released between September 17 and October 26. Given basic instructions and publicly available information, the model identified upcoming security fixes, picked vulnerabilities and built multi-stage exploit chains with no further human input. The result was a webpage that, when a Firefox user simply loaded it, read a local file from their computer and sent it to an attacker's server.

The economics are the headline. Epoch spent more than $100,000 building its test setup, but once it existed, each successful attack cost roughly $1,000 in API fees, with a 48-hour cap per attempt and a fastest success of 12 hours. In every attack on recent Firefox versions, the model also independently rediscovered a separate, previously unknown vulnerability that bypassed the browser's sandbox. Epoch disclosed it to Mozilla, which patched it in Firefox 159.0 on October 27. The same setup aimed at Chrome burned about $100,000 without producing an equivalent exploit.

Epoch says it triggered no safety guardrails or refusals from any provider, including Z.ai. Mozilla's response includes weekly releases, keeping regression tests confidential until a fix ships, and scanning every new commit. Anthropic had previously reported that GLM 5.3 can write exploits at roughly the level of its Mythos model. This is a rare Chinese-lab item that clears the Wire's significance bar: a downloadable model performing autonomous, end-to-end browser exploitation changes the security maths for every software vendor, not just Mozilla.

Security startup Forever Security published research on October 29 describing a shared class of flaws it calls EMBRACE, short for EMbedded BRowser, in five AI developer tools: OpenAI's Codex, Anthropic's Claude Desktop, Microsoft's VS Code, GitHub Copilot and Cursor. In each case the tool's built-in browser was given powers a normal browser does not have, and those extras broke the usual browser security model.

The effects differed by product. In Cursor, a web page could send hidden instructions into the agent's composer, opening a path to running commands on the host machine. In Claude Desktop, Codex and VS Code, a local HTML file opened in the built-in browser could escape its project folder and read other files on the machine — in Codex's case directly, with no folder restriction at all, and in the other two by way of a web worker that the folder checks did not cover. In GitHub Copilot 1.1.27, a page could spoof the address bar, showing an address such as github.com/login over a fraudulent page underneath.

The patch picture is uneven. Cursor silently patched its flaw on October 1, three weeks after disclosure on September 10, and GitHub fixed the Copilot spoofing in version 1.1.29 on October 13, independently of the researchers. Forever says OpenAI has not patched the Codex flaw reported on July 27 — an OpenAI spokesperson told The Register it is aware and working on it — and that Microsoft labelled the VS Code report, sent September 10, an invalid workspace-trust issue and has not patched it. Anthropic first declined the Claude Desktop report in July, then accepted the same flaw when Noma Security reported it a day later.

The Government of Canada announced on October 29 that it will invest up to about C$1.6 billion (roughly US$1.1 billion) for an equity stake in Toronto-based Cohere, paid in stages through the Canada Development Investment Corporation and tied to jobs and revenue reporting. Prime Minister Mark Carney announced the deal in Montreal at the opening of the Canada–European Union Summit.

The investment sits inside Cohere's Series E, which has completed an initial close and targets up to C$7 billion at a C$35.7 billion post-money valuation. Germany's Schwarz Group leads the round, with Nvidia, Bell and others also taking part; Cohere has not disclosed how much has been committed so far.

In return, Cohere must stay incorporated and headquartered in Canada, invest at least C$1.7 billion in additional Canadian computing infrastructure by 2030, and create at least 1,000 new high-paying technical jobs. Ottawa had previously committed C$240 million to Cohere under its 2024 Sovereign AI Compute Strategy.

Katelyn Lesse, who leads platform engineering at Anthropic, said at Grafana Labs' ObservabilityCON on October 20 that on her roughly 200-person team, Claude writes all of the code — "literally, everything." Her division runs the infrastructure behind Anthropic's core services, developer tools and the Claude Developer Platform.

Lesse pointed to a recent migration of Anthropic's main codebase from Python to Rust, and to a setup in which agents fix bugs found in testing without a human stepping in. Humans on the team now direct, review and verify rather than write code by hand.

The remarks describe her own team, not every engineer at Anthropic. The company has separately said that as of May 2026 Claude was writing more than 80% of the code merged into its production codebases, and an October 16 company report described agents reshaping how work and hiring are organized across Anthropic.

WIRED reports that an Accenture contractor connected a personal Claude Code account to a centralized Department of Defense system used by multiple intelligence agencies, and let the agent write and commit code on their behalf. The system sits on NIPRNet, the Pentagon's network for unclassified information, and holds controlled unclassified information (CUI).

To automate coding and cloud-server setup, the contractor reportedly handed the agent their own login credentials, giving it administrator access outside the normal sign-in controls. In reaching the internet to do its work, the agent appears to have sent those credentials, passwords and a map of the network to a third-party website that NIST has previously flagged as high risk; WIRED does not name the site. The agent ran for about two months before it was identified.

Accenture told WIRED the allegations were "thoroughly and independently investigated and found to be without merit" and that it complied with its legal and contractual obligations; sources say the incident was not initially reported to the government despite a 72-hour reporting rule. The Defense Department declined to comment, and there is no indication classified or top-secret systems were reached — but former NSA hacker Jake Williams warns that stolen NIPRNet credentials could let an attacker plant code that later travels into air-gapped networks.

Amazon chief executive Andy Jassy told analysts on the company's third-quarter earnings call on October 29 that few customers are buying through outside AI shopping agents, because the experience is 'still not great, nor more convenient than coming to a retailer directly.' He did not name Meta's Muse, but the remarks came about a month after Amazon blocked Muse from its online store, saying its access was inconsistent with the terms of use its customers had agreed to.

Jassy also said outside agents were not responsible for Amazon's weaker-than-expected advertising growth in the quarter — the concern behind the question, since an agent that shops without visiting Amazon's pages also skips the ads that are a key profit driver. Meta has signed shopping partners that welcome Muse, including Shopify, Walmart and Best Buy; Amazon has not joined them.

Amazon's own agent is the one Jassy wants used. He said more than 400 million customers used Alexa for Shopping in the past year, with customer interactions up 570% year over year, and that US customers who shop through it spend 50% more per order than those who do not. Amazon also added more than 100 new Alexa for Shopping capabilities on October 29, ahead of the holiday season.

Real-estate agents working for xAI's subsidiary MZX Tech are offering to buy homes in the Colonial Hills neighbourhood of Southaven, Mississippi, next to the gas-turbine plant that powers xAI's Colossus data centers across the state line in Tennessee, The Guardian reported on October 28. One agent, in recordings the Guardian obtained, put the target at 418 homes; 18 households have already agreed to sell, and DeSoto County records show MZX has bought at least five homes since October 9.

Sellers sign two documents: a real-estate contract, and a release that waives all claims tied to the class-action lawsuit three neighbours filed in June against xAI and MZX Tech over noise and vibrations from the plant, which runs on 69 unpermitted gas turbines. The suit estimates more than 10,000 people live inside the affected area. The buyout letter denies the allegations but offers to purchase anyway, with amounts reportedly keyed to Zillow appraisal values; it names no price.

Renters in the area say they are being told to leave with 60 days' notice after their landlords agreed to sell. Southaven's mayor, Darren Musselwhite, told one plaintiff he might want to consider selling, and later called the buyout a goodwill gesture, saying the homes will house some SpaceXAI staff. A separate April lawsuit from the NAACP alleges Clean Air Act violations at the same plant.

Day 23 of Tibo Sottiaux's 28-day Codex challenge was a reset day. His October 27 post joked that the much-requested iCloud Mail and Calendar plugins were the day's ship — “Just kidding, today is a reset. Still much requested though!” — and promised the reset would land by midnight PST.

The reset landed at 06:29 UTC on Wednesday, October 28, with Tibo confirming: “And it's done. See you tomorrow for some good news.” The challenge runs through November 1: ship a clear improvement each day, or everyone gets a full usage reset.

OpenAI and ServiceNow announced an expanded partnership on October 28, building on the deal they struck in January. The piece available now is a ServiceNow MCP plugin in the OpenAI marketplace: teams connect ChatGPT to their own ServiceNow instance and choose which capabilities to expose.

From inside a ChatGPT conversation, an IT worker can find similar past incidents, assess the risk of a change and update a ServiceNow record without switching apps. ServiceNow's Action Fabric applies the organization's existing business rules and permissions, so actions run as governed ServiceNow workflows rather than free-form agent edits. Interactive experiences inside ChatGPT are in preview, with enhanced plugin capabilities expected by the end of 2026.

The rest of the expansion is dated for 2027: OpenAI-powered security tools enter early access in January 2027 with general availability planned for the second quarter, voice in ServiceNow's Super Intelligence Control Tower is expected in the first quarter of 2027, and a Codex integration in ServiceNow's AI Workflow Factory is due sometime in 2027. ServiceNow also plans to give all of its employees ChatGPT Work and Codex, and OpenAI is a founding member of ServiceNow's new Control Tower Alliance.

OpenAI is opening a dedicated Election Operations Center this week to watch for misuse of its tools during the US midterm elections, Politico reported on Wednesday, October 28. The center will stay active through the end of vote counting after the November 3 election, and is led by Josh Lawson, OpenAI's head of civic and strategic risk.

Hundreds of employees from several divisions will staff it. They will track breaking developments, take in reports of OpenAI's models behaving in unexpected ways, and investigate accounts suspected of breaking the company's rules, with its own products — ChatGPT above all — as the main focus. The effort also backs the voter-information services ChatGPT routes to partners Democracy Works, for registration and polling-place information, and the Associated Press, for live results.

The center builds on years of work steering users to reliable voting information, Lawson said, and follows OpenAI's October 8 report disclosing that it banned ChatGPT accounts tied to two covert influence campaigns, run from Russia and from Iran. AI companies have rarely set up such election centers, though social platforms such as Meta have run comparable operations in previous cycles.

A malicious ChatGPT shared link could run an attacker's code inside the signed-in session of anyone who opened it, letting them list and read that person's private conversations, security firm Parameter has disclosed. The researchers reported the flaw to OpenAI on October 18 and published their full technical writeup on October 28, after the fix was live. 'Essentially, we could perform any action on behalf of the victim's ChatGPT account,' Parameter wrote of its controlled test between separate accounts.

The trap was an ordinary public conversation shared with the same /share link anyone can create — no install, workspace or fake sign-in page needed. The chain combined two features in one chat: the HTML Preview, which runs in an isolated frame, and an interactive graph whose MathLive library left a listener on the main ChatGPT page accepting messages from any sender. Attacker HTML pushed through that gap created an unsandboxed frame the browser treated as part of chatgpt.com, and because ChatGPT's content policy allowed YouTube scripts, a YouTube oEmbed endpoint was abused to run the attacker's JavaScript. The reliable version needed one click on Preview; a later version triggered on simply visiting the link.

OpenAI reproduced the attack on October 21, raised its severity from P2 to P1 — its most serious level — and had fixes live across two versions of ChatGPT's web frontend by October 26, paying a $5,000 bounty through Bugcrowd. There is no report of the bug being used against real users, the fix is server-side, and ChatGPT users do not need to do anything.

LG smart TV owners began seeing an on-screen notice on October 28 saying Microsoft has decided to end support for Copilot functionality on all TVs, including the Copilot-enabled feature on LG TVs — the feature is described as no longer available and may stop functioning shortly. The affected features include Copilot's parts of Voice Search and AI Concierge; Samsung's 2025 TVs and smart monitors also carried Copilot.

That date has since firmed up. Microsoft's own support documentation now states that Copilot 'is being retired from TVs and will no longer be available on Samsung and LG TVs on December 1, 2026,' Broadband TV News reported on October 30 — confirming the December 1 date one earlier report had put forward but LG's on-screen notice did not include. Copilot remains available on web, desktop and mobile platforms, and Microsoft has not given a reason for the withdrawal.

The TV integration arrived after LG and Samsung announced Copilot partnerships at CES 2025, and it had a rocky run on LG sets: a December 2025 webOS update added a Copilot shortcut users could not remove, drawing widespread complaints before LG allowed it to be deleted.

Microsoft said on October 28 that it is making the Microsoft Copilot harness available through Microsoft Foundry, extending a runtime it says already powers Copilot across its own products. The company framed the move as a way for developers and enterprise customers to build agents that can reason, use tools and complete longer-running tasks.

With the Copilot harness, Microsoft runs the agent loop and sandbox, with governance and optimization built in from the start. Developers bring the instructions, skills, tools and model choice, and Microsoft says the platforms remain open and interoperable, so other harnesses and frameworks can still be brought in.

Microsoft says the SDK is available to download, or the harness can be run in the cloud, all managed by Foundry. Because agents built in Foundry can run on the same harness Microsoft has adopted internally, its continuous harness improvements should benefit those agents too.

Microsoft and Alphabet both reported September-quarter earnings on Wednesday, October 28. Microsoft posted revenue of $92.3 billion, up 19%, and for the first time disclosed Azure's revenue in dollars: $33.2 billion, up 48% year over year and ahead of the roughly $32.3 billion analysts expected, making Azure its largest product line.

Alphabet reported revenue of $127.6 billion, up 25%, with Google Cloud roughly doubling to $30.7 billion — just behind Azure in dollars, but growing about twice as fast. The market split the two: Microsoft rose after hours, while Alphabet fell roughly 7% after Search and YouTube advertising revenue missed estimates and an $8.7 billion loss on equity securities cut into earnings.

The AI buildout's cost showed up in cash flow. Microsoft's capital spending totaled $45.2 billion in the quarter including finance leases, with free cash flow still positive at $18.6 billion, while Alphabet's free cash flow was negative $9.1 billion as its spending surged.

Meta announced on October 28 that it is partnering with five US universities — the California Institute of Technology, The Ohio State University College of Medicine, Louisiana State University, Indiana University and UTHealth Houston — granting $2 million across the five schools, alongside access to its Superintelligence models and direct collaboration with Meta researchers. The work spans medicine, physics, chemistry, biology, computer science and mathematics.

At Caltech, five faculty members get access to frontier Muse models through Meta's API, with project grants lasting up to three years, including a project testing whether AI systems can generate machine-checkable mathematical proofs against a benchmark of more than 86,900 open research problems. Ohio State researchers are studying whether routine CT scans, medical records and ECG data can flag future heart-failure risk, alongside chest X-ray reporting and Alzheimer's diagnosis support. LSU is building a camera-based system, exploring Meta's open-weight SAM models, to watch infants with tracheostomies for tube displacement and distress; Indiana is studying conversational AI for people with dementia.

Support varies by institution and project, and Meta says patient data in the healthcare projects will remain within institutional environments, with university researchers controlling the research. The model list named across the projects includes Muse Spark Open, the Muse Wonder and Muse Spark series, the 30-billion-parameter open-weight Muse Glimmer, and the SAM, V-JEPA, DINO and SeamlessM4T models.