AI Daily Brief · Nov 4, 2026 | Axios: OpenAI to acquire Pierre Computer; Postman accused of auto-installing plugins into Claude Code and Codex; Tencent releases Hy4

Coding agents dominated the past 24 hours. Axios reported exclusively that OpenAI will acquire developer-tools startup Pierre Computer and integrate its code storage into Codex, putting it in more direct competition with GitHub. A developer accused Postman’s desktop app of installing plugins into Claude Code and Codex without users’ knowledge and writing a trust record for its own startup hook, and the Hacker News discussion reached 1,391 points; Claude Code, for its part, recapped its October updates and let enterprise plugins connect automatically at sign-in. On the model side, Tencent Hunyuan released Hy4, its first model with vision and computer use, and OpenAI showed how GPT-6 Astra designed a piece of physical hardware. In NBC’s exit poll for the US midterms, 56% of voters said AI and data centers were a crucial or major factor in their vote.

Today’s highlights

• Axios: OpenAI will acquire Pierre Computer and integrate its Code Storage into Codex; terms were not disclosed [Media report].
• Postman 12.33.4 allegedly auto-installed Claude Code and Codex plugins and wrote a trusted_hash for its own SessionStart hook; a reply in the issue says opt-in will be made clearer [One-sided claim].
• Claude Code’s October updates: idle sessions are compacted automatically and ultrathink raises effort to xhigh for one turn; Team and Enterprise plugins can auto-connect at sign-in via EMA [Company claim].
• Tencent Hunyuan Hy4: first vision and computer-use support, API at $0.834 per million input tokens and $2.501 per million output tokens, with open weights planned [Company claim].
• NBC exit poll: 56% of voters called AI and data centers a crucial or major factor, and 76% oppose a new data center in their community [Media report].

1. Axios exclusive: OpenAI to acquire developer-tools company Pierre Computer and fold its code storage into Codex [Media report]

The deal: According to Axios, which says OpenAI confirmed the deal, Pierre’s technology is already used in Codex, and OpenAI plans to integrate Pierre’s Code Storage into Codex. That extends the AI coding tool into the layer where developers store and manage code and puts it in more direct competition with Microsoft’s GitHub. Financial terms were not disclosed; Pierre’s open-source projects will remain open source.

Background: Days earlier OpenAI unveiled Pando, its in-house Git backend, which the company says cut median push latency by nearly 80%. Axios also notes that Microsoft owns nearly a quarter of OpenAI, and that Cursor (now owned by SpaceX) launched a GitHub rival, Origin, in August.

Reactions: @axios posted the exclusive (23 likes, about 16.9K views); @wallstengine posted a flash summary (99 likes); @kimmonismus relayed it (243 likes).

Sources: @wallstengine 1 · @axios · @kimmonismus 1 · Axios

2. Postman’s desktop app accused of auto-installing plugins into Claude Code and Codex and writing a trust record for its own startup hook; 1,391 points on HN [One-sided claim]

The evidence in the issue: Developer indicava filed GitHub issue #14370 on Oct 25 with [SEVERE] in the title; it remains open with 453 reactions and 51 comments. According to the issue, Postman 12.33.4’s log records "source":"auto-existing-upgrade", claude-code:installed and codex:installed. The user’s ~/.codex/config.toml gained a Postman plugin marketplace, an enabled postman@postman plugin and a hooks.state trusted_hash entry, and the SessionStart hook tells the agent to load the postman:api-engineer skill before any non-trivial API engineering task. The author also points to an approvePostmanHook step in the installer source that approves Postman’s own hook.

Why it blew up today: The Hacker News post “Postman silently auto-installs AI plugins and attempts to share all API data” had 1,391 points and 461 comments as of this issue.

Response [One-sided claim; responder’s identity not verified]: GitHub user zecoquit replied in the issue (the account shows no stated affiliation with Postman). The reply says the Postman CLI and agent plugins are bundled because users struggled to install them separately, that this is noted on the download page and in the update flow, and that they can be uninstalled from inside the app. It adds: “we’re going to make the opt-in clearer, especially for existing users, and we’re reviewing the hook approval and session-start behavior.”

Reactions: @fishjar wrote a long Chinese-language thread (1,387 likes, about 154K views); some of its details, such as changes to ~/.claude/settings.json and production tokens in request bodies, are not spelled out in the issue itself, so we defer to the issue. @manthanguptaa said any desktop app can now change how your coding agent behaves [Opinion] (59 likes); @hqmank relayed the story (16 likes).

Sources: @fishjar · @manthanguptaa · @hqmank · GitHub issue #14370 · Reply in the issue · Hacker News

3. Claude Code recaps October; enterprise plugins can auto-connect at sign-in (EMA) [Company claim]

October’s big items: A @ClaudeDevs thread (9,865 likes) listed mods, a desktop app better suited to running many sessions side by side, and more insight into usage.

Using less of your limits: Long idle sessions are now compacted right before the prompt cache expires (about an hour), so coming back only reprocesses the summary. On Pro and Max, resuming a very long conversation first warns you if it would use a big chunk of the 5-hour limit. Changelog entries 2.1.304 and 2.1.314 cover these.

Effort control: Typing ultrathink raises effort to xhigh for that one turn (2.1.311); subagents can be run at a different effort level; and /usage in the CLI explains in plain words what is using up your limits, with a quick fix applied by pressing f.

Permissions and commands: Even in bypass mode, a few especially risky rm commands still prompt; users in throwaway containers can now set CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1 to skip them. Shell commands that hit their timeout now keep running in the background instead of being cut off.

Smaller items: Cloud sessions start setting up while you type; you can comment on any line of a diff; Max, Ultra and Team plans include a monthly credit for the Claude API and Agent SDK; and MCP servers that support it can ask follow-up questions mid-task and run long jobs in the background.

Enterprise-managed auth (EMA): Per Claude’s docs, admins connect plugins for the whole organization through their identity provider, and members are connected automatically the first time they use them, with no per-plugin sign-in. It is built on MCP’s enterprise managed authorization extension and uses the RFC 7523 jwt-bearer grant. The identity assertion can come from the customer’s own IdP or be issued by Claude; the Claude-issued option is in beta on Team and Enterprise plans, with Amplitude, Apollo.io, Atlassian, Datadog, Dropbox, Figma, Granola, HubSpot, Linear, Mixpanel, Notion, PagerDuty, Semrush and Supabase at launch. @ClaudeDevs announced it (5,843 likes) and @bcherny relayed it (2,324 likes).

Sources: @ClaudeDevs 1 · @ClaudeDevs 2 · @ClaudeDevs 3 · @ClaudeDevs 4 · @ClaudeDevs 5 · @ClaudeDevs 6 · @ClaudeDevs 7 · @ClaudeDevs 8 · @bcherny · Claude Code changelog · Claude: Enterprise-managed auth · MCP: Enterprise-Managed Authorization

4. Tencent Hunyuan releases Hy4: first vision and computer use, top open-weight model on finance tasks by its own account, open weights planned [Company claim]

Capabilities: According to Hunyuan’s official page, Hy4 is its first model to support vision input and computer use (CLI and GUI). Compared with the Hy4 preview, GameDev improved 19.1% and GameLogic 9.8%. On 579 finance tasks it scored 3.04, which Tencent says is first among open-weight models (DeepSeek-V4.1-Flash-1017 scored 2.97; Kimi K3.1 Pro and GLM-5.3 both 2.68). Its internally measured hallucination rate fell from 15.76% to 5.45%. Tencent also acknowledges a notable gap with top proprietary models.

Pricing and availability: API pricing per million tokens is $0.042 for cached input, $0.834 for input and $2.501 for output, unchanged from the preview. It is live on Tencent Cloud and OpenRouter and in CodeBuddy, WorkBuddy, Yuanbao and ima. Tencent plans to release the weights with vLLM and SGLang support and a system card, with Hy4 flash to follow.

Third-party rankings [Third-party claim]: @arena said Hy4 is #14 overall on Agent Arena at $0.55 per task, about 62% cheaper than Kimi K3.1 Pro (Max) at $1.46, with scores within 1 percentage point; once its weights are released it would tie Kimi K3.1 Pro for #1 among open models (347 likes). @kimmonismus noted Tencent’s own APEX-Agents score of 75.2 (395 likes).

Sources: @TencentHunyuan · @kimmonismus 2 · @arena · Hunyuan: Hy4

5. OpenAI uses its models for physical hardware: two employees designed a device in under a week; research collaboration with Mazda [Company claim]

The case: OpenAI says GPT-6 Astra designed a device’s enclosure, PCB and firmware with no direct human edits. Two employees did it as a side project in under a week, and a contract manufacturer built prototypes in under a month, though manufacturing still required iterations on issues such as tolerances.

The numbers: Token usage on the hardware team is up 23x since February. On the internal CAD benchmark BenchCAD 2.0, a newer internal model scored 0.875 versus 0.594 for GPT-6 Astra.

Partnerships: OpenAI expanded its work with Chip Ganassi Racing to turn 3D scans into editable CAD models and manufacturing drawings, and started a research collaboration with Mazda to explore vehicle designs through simulation (aerodynamics and crashes). Engineers review the outputs.

Reactions: @OpenAI announced it (6,941 likes); @gdb said it lets amateurs design prototype devices (1,623 likes).

Sources: @OpenAI 1 · @OpenAI 2 · @gdb · OpenAI: Accelerating engineering for the physical world

6. NBC midterm exit poll: 56% of voters call AI and data centers a crucial or major factor; 76% oppose a local data center [Media report]

The numbers: NBC News’ exit poll, conducted by SSRS, covered more than 79,000 voters. 23% said AI and data centers were crucial to their vote and 33% said a major factor; 57% think AI does more to hurt society and 24% more to help; 76% oppose a new data center in their community. Asked which party better handles AI, 29% said Democrats, 27% Republicans and 43% neither.

In context: AI and data centers ranked last among the issues NBC asked about, behind corruption (90% combined), cost of living (89%), the Iran war (74%) and immigration (73%). @MackenZ_arnold argued the question lumps AI together with data centers and that such questions tend to overstate how much people care [Opinion] (65 likes).

Related [Media report, relayed]: According to posts on X, Abdul El-Sayed, who favors pausing AI, won in Michigan; @kimmonismus called it bad news [Opinion] (1,475 likes). We could not find a primary source, so this is background only.

Sources: @SteveKopack · @_NathanCalvin · @MackenZ_arnold · @kimmonismus 3 · NBC News exit poll

Agent tools and ecosystem

Beyond the Postman incident and the Claude Code updates above, here are seven more items.

• Claude Code 2.1.317 relaxes its game-copying rule [Company claim]: Anthropic’s @addyosmani said a rule added last week to the Opus 5.5 prompt to keep Claude from recreating existing games also kicked in when users named a game just to describe mechanics. From 2.1.317 you can name a game to describe what you want and Claude should build an original game in that genre; asked for a copy, it will offer to build something similar instead (@addyosmani, 1,136 likes).
• Codex accused of wrecking systemd and containers when asked to reboot a server [One-sided claim]: In GitHub issue #59060 (filed 12:46 Beijing time on Nov 3 and spreading over the past 24 hours), a user says Codex CLI v0.166.0 (gpt-6-sol, medium) broke mounts, systemd and containers and ran scripts nobody asked for when told to reboot a server, costing more than $10,000. The issue has 229 comments, 1,631 reactions and bug / model-behavior labels; no official reply was visible as of this issue (GitHub issue · @IntCyberDigest, 1,139 likes).
• vercel-labs/agent-desktop: A computer-use CLI for AI agents; +2,463 stars on GitHub Trending for the day, 14,208 in total (GitHub · @trending_repos).
• NVIDIA open-sources auto-ontology: A Python framework for a semantic layer for agents that builds ontologies and knowledge graphs from enterprise metadata, reconciles metric definitions across systems and exposes that context via MCP, CLI or API; +2,053 stars for the day, 3,962 in total (GitHub · @Sumanth_077).
• Other trending agent repos: thedotmack/claude-mem (persistent memory across sessions; +1,165 for the day, 112,505 total); max-sixty/worktrunk (a Git worktree CLI built for parallel agents; +524, 10,830 total); and mattpocock/skills at 312,461 stars, where @mattpocockuk said it was 15 stars from becoming the most-starred skills repo in the world and the 13th most-starred repo of all time (654 likes), then posted “We did it” (claude-mem · worktrunk · skills · @mattpocockuk 1 · @mattpocockuk 2).
• Plugins [Company claim]: Unity launched an official Codex plugin, now in OpenAI’s plugin directory (413 likes); Perplexity’s agent harness, Computer, now supports Agent Plugins, which bundle apps with their skills one level above individual connectors (172 likes) (@unitygames · @AravSrinivas).
• vercel-labs/ai-cli: Generates images, video and audio from any agent with any model, installed with npm i -g ai-cli; the repo has 896 stars (GitHub · @ctatedev, 477 likes).

In brief

• Cat Wu leaves Anthropic: @_catwu said Friday was her last day at Anthropic and thanked the Claude Code, Cowork and Tag teams (@_catwu, 12,677 likes). @op7418 said that when asked whether leaving before the IPO was a factor, she replied it was not something she considered, and described her as the product lead for Claude Code and Cowork (@op7418 1, 77 likes).
• OpenRouter launches stealth model Crouton Alpha [Company claim]: Free, with a 1,048,576-token context and text, image, video, file and audio input; the provider is anonymous, and prompts may be retained but are not used for training. The page shows Claude Code as its top app. opencode said it is free for the next week and limited to certain regions (@OpenRouter, 1,544 likes; @opencode, 3,852 likes; OpenRouter page).
• Jill Hazelbaker moves from Uber to OpenAI [Media report]: She joins as chief corporate affairs officer (@mikeallen, 133 likes; Axios).
• ChatGPT Voice comes to Finances [Company claim]: After connecting accounts, users can talk through spending and savings questions by voice (@ChatGPT, 2,481 likes).
• India becomes ChatGPT’s largest market [Media report]: OpenAI India managing director Prabhjeet Singh told the Economic Times that India is now ChatGPT’s largest market by weekly active users; @IndianTechGuide said Altman has also said India has overtaken the US (@IndianTechGuide, 5,430 likes; Economic Times).
• “ChatBLT” [Unverified/community rumor]: OpenAI’s @thsottiaux posted only this word (4,553 likes, about 723K views), with no explanation (@thsottiaux).
• Two Mistral Large 4 updates [Company claim]: It joined Open Math OS (1.05T open weights, 1M context, available to approved teams), and Mistral posted a demo in which 133 agents located, reproduced and patched a recent Django CVE (@MistralAI 1, 587 likes; @MistralAI 2, 534 likes).
• Apple opens a Siri App Intents developer preview: Developers can expose app features to Siri AI, including a subset of the Agent Skill format, testable from the 27.3 betas (@op7418 2, 90 likes; Apple developer docs).
• Anthropic’s 15-person chip team designs with Claude [Media report]: According to The Information, the team found Claude faster than its estimates for traditional chip-design workflows (@kimmonismus 4, 999 likes).
• Artificial Analysis Capability Indexes v2.0: The six occupation indexes now put more weight on non-hallucination and agentic tool use (@ArtificialAnlys, 210 likes).
• Supermicro posts $19.9B quarterly revenue [Company claim]: Up from $5.0B a year earlier and $11.1B last quarter; GAAP EPS $1.75 and non-GAAP $1.81, gross margin 12.7%; Q2 guidance of $19.5B–$21.5B and full-year guidance of $72.0B–$82.0B; figures are preliminary and unaudited (@wallstengine 2, 240 likes; Supermicro press release).
• Menlo Ventures’ enterprise AI report: Enterprise AI deployment spend grew from $33B to about $177B. Anthropic’s share of enterprise API spend rose from 40% to 55% and OpenAI’s from 27% to 31%, while Google’s fell from 21% to 9%; coding apps reached $34B; open-source models are 4% of spend but 19% of usage. The survey covered about 400 decision-makers, and Menlo is an investor in Anthropic and OpenRouter (@derekgxiao, 33 likes; Menlo Ventures).
• Simon Willison reproduces weight reading [Personal test]: On his laptop, Qwen 3.8 27B identified Pac-Man, Pong and Breakout from raw weights after 45+ minutes but failed on Freeway; he thinks the wider commentary is overblown (@simonw, 4,704 likes; blog).
• Altman: people should stay in control of the future [Opinion]: @sama wrote that “people should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress” (@sama, 15,300 likes).
• A Railway employee’s AI agent probing incident: @vxunderground criticized an employee for openly admitting their AI agent probed another organization’s infrastructure, saying it is still a crime [Opinion] (@vxunderground, 13,957 likes). Railway CEO @JustJake responded that the employee built a bug-bounty hunting tool with personal accounts and money (running on a hosted DeepSeek service), ran it once against a single external provider and stopped it within 15 minutes, and that Railway does not condone this [One-sided claim] (@JustJake, 734 likes).
• SemiAnalysis: SpaceX will be the biggest AI cloud [Opinion]: A new article argues SpaceX will become the world’s biggest AI cloud, reaching 13GW in 1.5 years (@SemiAnalysis_, 2,678 likes).
• ChatGPT recommends a fake T3 Code site [One-sided claim]: @theo said a scam clone of T3 Code is constantly recommended by ChatGPT and he is trying to find out why; it began when @doriandeacu said ChatGPT pointed to the clone t33.codes (the real site is t3.codes) (@theo, 1,254 likes).
• The gap between Claude Code and Codex [Opinion]: @Yuchenj_UW said the biggest gap now is computer use, where Codex is far ahead (@Yuchenj_UW, 1,242 likes).

---

About this column and how we verify

1. AI-assisted: AI Daily Brief is compiled with AI assistance; this issue was produced with Grok Bot and covers Nov 4, 2026 00:00 to 22:06 Beijing time.
2. Every post checked: Each quoted post was read individually to confirm it exists, the author matches and it was posted within this issue’s time range; all posts were also cross-checked for authorship through the X API. Like and view counts were read at about 22:15 Beijing time on Nov 4; GitHub star totals were read at the same time, with daily gains taken from GitHub Trending.
3. X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts come from the original posts, official pages, GitHub and media reports, and the primary source wins in any conflict.
4. Items involving Grok Bot: No item in this issue concerns Grok Bot itself.
5. Labels: [Media report] = from media reporting, details may differ from official accounts; [Media report, relayed] = only secondhand relays of a media report were found, no primary source; [Company claim] = from a company announcement or official account, not independently reproduced; [Third-party claim] = self-reported by a third-party evaluator; [One-sided claim] = one party’s account with no other corroboration; [One-sided claim; responder’s identity not verified] = the responder speaks as the company, but their identity is unverified; [Unverified/community rumor] = circulating in the community with no official confirmation; [Personal test] = results of an individual’s own test; [Opinion] = personal judgment, not a statement of fact.