AI news daily brief: 2026-11-02
Six stories from November 2: a security incident, a model-pricing milestone, a breaking framework update, and three GitHub product moves that all trace back to the same October 28 Universe wave, five days old by the time this lands. One of the three, Agent Merge, already has a published Pondero guide worth a second look before a team pairs it with auto-merge. The day's biggest story, an Anthropic model that chained 62 Chrome CVEs into five working exploit chains across Chrome, Safari, Android, Windows, and Linux, got its own deep-dive at the Anthropic exploit-chains piece.
GitHub's Agent Merge has been watching pull requests and auto-fixing CI since October 28
GitHub built Agent Merge into the pull request page on GitHub.com starting October 28, 2026: Copilot watches an open PR and fixes the two things that usually stall it, a failing Actions workflow or a merge conflict, without a re-prompt, per GitHub's changelog. It needs a paid Copilot plan (Pro starts at $10 a month) and skips forks and stacked PR chains. Pondero's guide to the feature, published the next day, walks through the branch-protection gap worth checking before pairing it with auto-merge: GitHub only dismisses stale approvals on a new push when a repo has turned on "Dismiss stale pull request approvals when new commits are pushed" in branch protection, an opt-in setting. Skip it, and a conflict-resolution push from Copilot can clear the way for a merge nobody re-reviewed.
n8n's 3.0 beta drops npm installs and cuts the Code node's timeout to 1 minute
Per n8n's breaking-changes changelog, the company shipped version 3.0 as a public beta on October 27, 2026, with a stable release targeted for early December. Self-hosted n8n now requires Docker: npm and npx n8n installs stop working, because 3.0 does not publish a runnable n8n package to npm at all, though Kubernetes and Helm deployments are unaffected. The release also tightens the Code node's default execution timeout from 5 minutes to 1 minute (N8N_RUNNERS_TASK_TIMEOUT drops from 300 to 60), configurable but no longer generous by default, and retires Chat Hub. A workflow leaning on a long-running Code node task needs that variable set explicitly before upgrading, or the task fails on its next run, a wrinkle Pondero's n8n self-hosted vs. cloud guide walks through in full.
GitHub's AI Scan can now crawl a whole repository, not just a pull request's changed lines
GitHub's changelog detailed the move of AI Scan for repositories into public preview on October 28, 2026, on github.com and GitHub Enterprise Cloud with data residency, a scope jump from the pull-request-only version it shipped in September. A scan analyzes the latest commit on a repository's default branch and flags issues in languages CodeQL has never supported. Public repositories get it without a Copilot license; private and internal repositories still need GitHub Code Security or Advanced Security, and usage draws down Code Security AI credits instead of a seat. Rescans do not auto-close fixed alerts, so a team must kick off a fresh scan after merging a fix to clear its own backlog. GitHub confirmed the same day that its narrower, pull-request-scoped sibling, "Code scanning for AI applications," stayed in public preview as a distinct product. Pondero's full write-up breaks down the credit math and the rescan gap: GitHub's AI Scan now explores whole repositories, not just pull request diffs.
OpenAI patched a ChatGPT share-link bug that ran attacker code in a victim's own session
Security firm Parameter reported a ChatGPT flaw through OpenAI's bug bounty program on October 18, 2026, detailing the find in a writeup published after the fix; OpenAI shipped that fix eight days later, on October 26. Opening an ordinary shared conversation link, one containing nothing more than an interactive graph and an HTML Preview block, was enough to run attacker JavaScript inside the victim's own logged-in chatgpt.com session: a math-rendering library loaded to support an on-screen keyboard accepted raw HTML through a postMessage listener that never checked which frame sent it, smuggling code past the sandboxed Preview frame's isolation. The code then called ChatGPT's authenticated APIs directly, listing and reading a victim's private conversations. OpenAI raised its internal severity rating from P2 to P1 and paid $5,000 across the two reports; Parameter found no evidence of exploitation against real users. Pondero's write-up has the full attack chain, frame by frame: OpenAI fixed a ChatGPT share link flaw that ran attacker code in a victim's signed-in session.
OpenAI's Decisions API reaches general availability at $0.10 per million input tokens
OpenAI's general-availability announcement said the Decisions API reached GA on October 22, 2026, priced at $0.10 per million input tokens with no charge for output tokens and a $0.01-per-million rate on cached input. Built on GPT-6 Luna, the API answers narrow, bounded questions (a probability, a choice, or a score) up to 10 times faster than calling Luna directly through the Responses API, and since beta OpenAI has added US and EU inference residency, native JSON and image inputs, and doubled the per-request question limit to 400. Named customers report concrete gains: Sierra said the API runs live customer-conversation routing at under 280 milliseconds p95, about three times faster and 90% cheaper per call than its prior approach, and language-learning platform Speak cut the time between AI actions from 1.4 seconds to 0.13 seconds. Pondero's GA write-up has the rest of the latency and cost comparisons.
GitHub open-sources Replica, an AI prototyping tool constrained to a team's own design system
GitHub's changelog entry announced that its design team released Replica as a free, MIT-licensed repository on October 28, 2026: it generates interactive mockups constrained to a team's own components and patterns rather than whatever a model invents. The repo ships a plugin that opens a live preview canvas inside the Copilot app, an optional self-hosted "Replica Hub" for publishing prototypes, and two starter kits, one built on Microsoft's CoreAI design system. Five days after launch, the repository had drawn 132 stars, 17 forks, and 72 commits from at least three contributors (GitHub's repo page has the running count), a small but real early-adoption signal with no outside customer case study yet. Pondero's take on the launch is here: GitHub open-sources Replica, an AI prototyping tool built from design systems.
Sources
- GitHub changelog: Agent Merge watches and fixes pull requests in the cloud: GitHub, October 28, 2026
- n8n changelog: v3.0 breaking changes: n8n, October 27, 2026
- GitHub changelog: AI Scan for repositories in public preview: GitHub, October 28, 2026
- Parameter: Do the Math, How We Took Over ChatGPT Accounts: Parameter, primary researcher writeup
- OpenAI: Build more responsive apps with the Decisions API: OpenAI, October 22, 2026
- GitHub changelog: Introducing Replica, an open-source AI prototyping tool: GitHub, October 28, 2026
