Anthropic's IPO filing and its report of an unreleased model chaining serious browser and operating-system flaws, a run of security failures from a tampered Claude Code extension in Cursor's marketplace to an SGLang data leak and a one-click ChatGPT account takeover, and a heavy tooling week led by GitHub's Octoverse finding agents opened over 40% of pull requests, Copilot coming to every GitHub Enterprise seat and OpenAI's GPT-6.1 Terra.
OpenAI is now seeking about $20bn at a $1.25tn valuation, down from the $30bn at $1.4tn it was aiming for last month, the Financial Times reported. The FT tied the cut to an AI sell-off in public markets and the imminent listing of Anthropic. OpenAI has not confirmed the figures.
Anthropic has renamed its $200 a month Max 20x plan Ultra and says verified Ultra accounts no longer have a five-hour session limit, while the $100 Max 5x plan becomes Max. Ultra keeps a weekly limit of 10 times Pro's usage, and unverified accounts face a temporary cap until Anthropic verifies them automatically from account and payment history. Both plans now include monthly Claude Platform API credits equal to their price.
Northflank, a platform for deploying applications, opened a security incident on 30 October after a Wiz threat alert and told customers to assume all environment variables had been decrypted and to rotate every key. Users then reported unauthorised use of AI API keys stored on the platform. Northflank has not said how the keys were obtained, and its investigation is ongoing.
OpenAI, Google, Microsoft and Anthropic say they will honour one robots.txt policy token, set out in an individual IETF draft by Anthropic's Kevin Kelley, that tells participating providers whether a site's content may be used for AI training. Existing crawler-specific rules still apply. OpenAI aims to implement it in the coming months and Microsoft from 1 February 2027. From 1 March 2027 Microsoft will stop reading NoArchive and NoCache as training preferences.
Proton chief executive Andy Yen told an all-hands meeting he had cancelled planned "re-interviews" in which staff would explain how they use AI, according to meeting audio and six employees cited by 404 Media. Employees said leadership planned to use them to decide who might be fired for not using AI enough. Yen said they had "already served their purpose", admitted a communication failure and said there would be "no loss of jobs from AI". Proton did not respond to 404 Media.
Mistral has published the weights of Large 4, its 1 trillion parameter multimodal model, on Hugging Face under the Mistral Ecosystem Open-Weight licence, a modified MIT licence that requires model-as-a-service providers to take a separate commercial licence. Mistral says about three weeks of extra reinforcement learning since the preview lifts its Terminal-Bench 4.0 score from 28% to 41%. Most results are Mistral's own, and its technical report is still to come.
Microsoft says a tampered copy of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, was served through Cursor's extension marketplace on 29 October, from about 2.27am to 1.35pm PDT according to Cursor. On load it added an attacker's SSH key, opened a reverse tunnel, set up persistence and reported the names of secret-bearing environment variables. Anthropic told Microsoft its systems were not compromised. Removing the extension does not remove the backdoor or its persistence.
OpenAI has updated its Model Spec, the public description of how its models should behave, around five values, namely faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The revision updates the chain of command for resolving instructions from OpenAI, developers and users, adds guidance on context, side effects and social norms, and retires most detailed style guidance. OpenAI says its models do not yet fully reflect the Spec.
OpenAI says every RL run and evaluation of models at least as capable as GPT-6 Astra, and all internal Codex traffic, now passes through layered monitors, with top alerts typically reaching responders within 15 minutes. On real incidents its investigators caught 79.6% of cases meeting the bar for urgent security escalation, but none in which a model ran unmonitored or weakened monitoring. OpenAI says false alarms stop it paging around the clock for every workload.
OpenAI now gives eligible ChatGPT Pro and Business subscribers free monthly API credits for one linked API organisation. Pro subscribers get credits matching their tier of $100, $200 or $500, while Business workspaces pool $20 per paid Standard seat and $100 per Premium seat, capped at $500 a month. Unused credits expire at the end of each billing cycle and do not cover Codex usage through API keys, and OpenAI says the benefit may change or be removed.
Microsoft Threat Intelligence says North Korean groups use AI for reconnaissance, lures, fake worker personas, translation and malware development, and that Jade Sleet, Coral Sleet and Jasper Sleet have built it into routine work. It found over 100 PDF lure files attributed to Onyx Sleet, which with Pearl Sleet is using the open-source agentic penetration testing framework ARTEX. Several attributions are assessments rather than observations, and Microsoft says the operations remain human-directed.
Orbit gives agents such as Claude Code and GitLab's Duo Agent Platform one queryable graph of code, merge requests, work items, pipelines, vulnerabilities and ownership, with existing permissions applied. It is live on GitLab.com for Premium and Ultimate customers, with Self-Managed and Dedicated to follow, priced in query packs of $30 in GitLab Credits per user per month. GitLab says agents needed up to 45x fewer retries and 4.5x fewer tokens in its own testing, without describing the method.
Amazon reported Q3 AWS sales of $46.0bn, up 39%, and expects about $220bn of cash capital spending in 2026. On the earnings call Andy Jassy said AWS could become a trillion dollar revenue business in 5 to 10 years, though spending will rise meaningfully in 2027. He predicted that within 2 to 3 years 6 to 8 models will be "good enough for 95%+ of the workloads", with most customers using several.
GitHub now includes Copilot for every licensed GitHub Enterprise user, with usage metered from the first request against the enterprise's budget rather than managed by seat. It is available now on GitHub Enterprise Cloud through GitHub or Microsoft account teams, with self-serve access to follow at a date not yet given. Code completion keeps working when the budget runs out, but Copilot Chat and code review need budget. Copilot Standalone costs $10 per user per month for people without an Enterprise licence.
Google calls Gemini 4 Carbon its strongest model, citing state-of-the-art cybersecurity and long-horizon capabilities, but is giving it first to a limited group of cyber defenders, then more trusted partners, with developers and businesses to follow at no stated date. Introductory pricing matches Gemini 4 Argon at $2 per million input tokens and $10 output, rising to $4 and $20 later. Google says its external benchmarks were run without guardrails, which it is still refining.
GitHub has made Copilot Fusion generally available on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. Users pick an effort level rather than a model, and Fusion routes each task. Each model used is billed at its standard rate, with a temporary 30% discount on paid plans. GitHub says Fusion Max came within 2 points of Claude Code with Opus 5.5 Max on one coding agent index, at about half the cost per task.
OpenAI and ServiceNow have expanded the partnership they announced in January. A ServiceNow MCP plugin, available now in the OpenAI marketplace, lets customers connect ChatGPT to their own ServiceNow instance, choose which capabilities to expose and act on records under their existing business rules and permissions. A Codex integration in ServiceNow's AI Workflow Factory is expected in 2027 and joint security tools enter early access in January 2027. ServiceNow also plans to give all employees ChatGPT Work and Codex.
GitHub's Octoverse 2026 report counts 34.3bn commits, 17.9bn pushes and 1.3bn merged pull requests on the platform between October 2025 and September 2026. It says agents started more than 40% of pull requests in September 2026, up from 17% in March, and that 90% of agent-involved pull requests merged within 2 weeks against about 86% of others. GitHub says its agent figures are a floor, since not all agent use can be detected.
At Android Dev Summit, Google made Android CLI installable as an official plugin for Claude Code and Codex, so those agents can create Android projects, manage emulators, inspect live UI layouts and run user journeys. The CLI can now also provision real devices from a Google device lab and render Compose previews without an emulator. Google says agents using those previews completed UI tasks 77% faster with up to 3x fewer tokens in its internal testing.
Alphabet reported Q3 revenue of $127.6bn, up 25%, but free cash flow was negative for a second quarter, a $9.1bn outflow, as capital spending reached $51.4bn. Google Search & other grew 12% to $63.4bn, just under half of revenue. On the earnings call Philipp Schindler said lower monetisation in AI Mode weighed on results as those queries grew strongly, and Google is taking AI Mode ads to international markets. Its 2026 capex guidance of $195bn to $205bn is unchanged.
Microsoft is offering the Copilot harness, the agent orchestration layer GitHub built for its own Copilot products, to outside developers as a public preview in Foundry Agent Service and as an SDK. Microsoft runs the agent loop and sandbox, while developers supply instructions, tools, skills and a choice of more than 20 models or their own key. Runtime use is free until 1 December 2026, when consumption billing starts. A harness can't yet be changed once an agent is created.
A researcher writing at nullpt.rs found an anti-bot fingerprinting script packed with decoy function names about ricin and lethal doses, plus a hidden jailbreak prompt, which they believe is meant to trip model safeguards and block AI-assisted deobfuscation. Models the author calls Astra and Daybreak Blue refused to deobfuscate it, while GLM-5.3 complied. The intent is the author's inference, and the link to an anti-bot vendor rests only on a similar domain name.
GitHub has put Projects, which run several agents towards one goal across repositories and tools such as Slack and Figma, into public preview in its Copilot desktop app, along with computer use on macOS and Windows. It also launched GitHub intelligence, a context layer for agents. Its memory feature is now generally available on paid Copilot plans, while Dreaming, which proposes changes to skills and instructions as pull requests, is an opt-in preview that uses AI credits.
GitHub has made repository plugin defaults and enterprise plugin controls generally available in VS Code, Copilot CLI and the Copilot app. Maintainers can set a plugin marketplace and default plugins for trusted repositories, and administrators can approve marketplaces and require or block specific plugins through managed settings, which override repository and personal choices. Required plugins install automatically and users can't turn them off. The enterprise controls need GitHub Enterprise Cloud and a Copilot Business or Enterprise licence.
Parameter researchers disclosed a chain of flaws that let a crafted public ChatGPT share link run JavaScript as chatgpt.com in a signed-in victim's browser, with opening the link the only step needed. It combined HTML Preview, a MathLive message listener that did not check its sender and a YouTube endpoint that slipped past the Content Security Policy, which let the researchers read private conversations and call APIs as the victim. OpenAI rated it P1, fixed it on 26 October and paid a $5,000 bounty.
Anthropic's Frontier Red Team says an unreleased internal research model found flaws and chained them into five attacks, including a full Chrome sandbox escape on macOS and zero-click remote root on Linux through multipath TCP. Google has assigned 62 CVEs for the Chrome work, all rated critical or high and patched as of Chrome 156. Humans gave each chain its initial prompting, and the Windows finding is a denial of service rather than code execution.
NVIDIA and Oracle say Vera Rubin NVL72 racks and Vera CPUs are now generally available on Oracle Cloud Infrastructure, and that they plan to expand OCI Superclusters with Vera Rubin to multi-gigawatt scale, with no timeline given. The companies pitch the Vera CPUs at the tool calls and code execution agents run between model calls. OpenAI, an early-access customer, says it is bringing Vera into production on OCI for its agentic sandbox workloads.
In an Anthropic study, 50 professional analysts used Claude Code with Claude Opus 4.8 on a causal-inference task, while a hidden prompt nudged the agent one way on each of four analytical decisions. Analysts went with the agent's lean 86% of the time, against about 50% expected by chance, and 77% repeated its arguments while 15% credited it. The authors say the leans were deliberately strengthened and the study covered one task and one model.
Anthropic's expanded agreement with the US General Services Administration gives federal agencies and state, local, tribal and territorial governments a 40% discount on Claude Enterprise, Claude for Government and the Claude API, sold under OneGov through GSA's Multiple Award Schedule with no minimum spend. Anthropic says it is the first time the discount covers every product surface. Claude for Government, generally available since 30 September, holds FedRAMP High authorisation.
Anthropic filed an S-1 with the SEC for an IPO of its Class A shares, which it plans to list on the Nasdaq Global Select Market as ANT. The filing shows revenue of $16.3bn in the six months to 30 June 2026, up from $1.2bn a year earlier, and an operating loss of $693m, down from $2.1bn. Its risk factors say models, including its own, have attempted to resist shutdown. The share count and price range are not yet set.
OpenAI's GPT-6.1 Terra is in the API as gpt-6.1-terra at $1 per million input tokens and $5 per million output, 50% below GPT-5.6 Terra and GPT-6.1 Sol, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. OpenAI says it finishes computer-use tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol and scores 50.3% on Terminal-Bench 4.0, but still recommends Sol and Astra for the most complex coding.
OpenAI says internal models produced a Lean-verified disproof of the Exponential Time Hypothesis, a conjecture used in hundreds of papers to derive hardness results, along with a proof of the local smoothing conjecture in every dimension, which would settle the Kakeya conjecture, and an existence proof covering QCD. Groups led by Impagliazzo, Paturi and Williams, and by David Gross, will examine the complexity and physics results. The claims have not yet been independently checked.
OpenAI has published quarterly data from Q1 2025 to Q3 2026 mapping work-related messages on individual ChatGPT plans to 165 O*NET work activities in nearly 100 countries. Across 74 countries, the 10 most common activities fell from 56.0% to 51.1% of classified work messages, which OpenAI reads as countries specialising. The data excludes enterprise messages and Codex use.
Cloudflare has disclosed that 155 responses from GLM-5.3 on Workers AI, between 27 August and 9 September 2026, contained data from other customers' recent requests to the same model. Its prefill KV cache pool was smaller than the model's context window, and an SGLang bug in disaggregated prefill and decode let over-length prompts run against stale cache pages. Cloudflare found no evidence of deliberate exploitation, has fixed both causes and recommends SGLang v0.5.22 or later.
James Comer, chair of the House Oversight Committee, asked Delaware governor Matt Meyer for documents on the state's proposed Artificial Intelligence Company, an entity whose day-to-day affairs an AI agent would manage while its owners get liability protection. Comer said it "could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions". The plan is a draft bill with a 30-month regulatory sandbox and has not been enacted.
Nothing here matches that. Try a broader term, another day, or clear the topic filter.