Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 2, 2026 · morning edition

The builder news going into Monday is rulebooks written for machines to read, a shared robots.txt training token Google has committed to and a DeepSeek agent that acts only on allowlists and written rules, while Microsoft's threat report shows attackers running agent frameworks with no rulebook at all.

Machine-Readable Training Preference SignalDeepSeek dsh-superagentTrigger.devMicrosoft Threat IntelligenceMOFCOM Announcement No. 48agent-securitymulti-agentregulationprivacyagent-harness

Trending AI Briefing: Monday, November 2, 2026 (morning ET)

No frontier model shipped over the weekend, and the quiet made something else easier to see. The interesting launches were rules that machines are supposed to read and obey: a two-line robots.txt group that Google has committed to honor for AI training (OpenAI reportedly too), and a DeepSeek research agent whose whole design is an allowlist plus team-written rules. Both depend on the machine choosing to comply. Microsoft's latest threat report is the reminder that the other side doesn't.

What's hottest in AI news right now

A shared robots.txt token for AI training now has a Google commitment, and reportedly OpenAI's. On October 30, Kevin Kelley of Anthropic published an individual IETF Internet-Draft, "Machine-Readable Training Preference Signal," that defines one policy token, __usage_ai-training. A site writes User-Agent: __usage_ai-training and Disallow: /, and any provider that honors the draft treats that as a preference against training on those paths. The draft adds no new grammar, says the token "is not a directive for crawling behavior," and says wildcard User-Agent: * groups "MUST NOT" be read as matching it. Google's crawler docs say it intends to support the token in early 2027 (John Mueller: "We aim to support this new control along with other AI providers in 2027") and will read an opt-out as if the site had disallowed Google-Extended, unless the site names Google-Extended explicitly. OpenAI's own post was unreachable from here; secondary coverage says OpenAI will implement it "in the coming months" and calls it voluntary. The catch: an individual draft has no formal IETF standing, nothing forces a scraper to read it, and it does not pull content out of training sets that already exist. IETF draft · Google crawler docs · Search Engine Roundtable · Unite.AI (secondary)

DeepSeek's dsh-superagent prototype is the top repo on Trendshift this morning. The deepseek-ai organization pushed a single commit on October 31 (Beijing time): a coding agent for one GitHub repository, packaged for DeepSeek Harness, that watches GitHub and optionally Feishu, turns an allowed person's request into a ticket, and answers in the same thread, often with a pull request. Its stated principle is to "take as little human attention as possible, and give as much value as possible." Proactive work needs a written team rule, and the default accept mode is observe, which records what the agent would have done instead of doing it. The catch is in its own docs: worker sessions "run commands without approval as the operator's OS user," and the directory wrappers "establish defaults, not a security boundary." The README also says it is "not an official DeepSeek product." GitHub · Execution doc

Trigger.dev disclosed that an attacker read customers' environment variables. The background-jobs platform posted a disclosure dated Sunday, November 1, saying "the attacker gained unauthorized access to environment variables" for affected organizations, and that its investigation found "exposure across a subset of organizations." Its status page marked the incident resolved at 22:50 UTC the same day and says the exploit is blocked. There is no root cause, no count of affected organizations, and no timeline of when access began; affected account owners got mitigation steps by email. If your jobs there hold API keys for model providers or databases, rotate them now rather than waiting for the follow-up. Trigger.dev disclosure · Status page

Microsoft's October 29 report on North Korean threat actors rates their overall AI adoption as "high," and the detail that matters for builders is the agent one. Onyx Sleet and Pearl Sleet are using ARTEX, an open-source agentic penetration-testing framework, in operations observed since late October 2026. The report also says Onyx Sleet produced more than 100 PDF lure files in August and September, "dozens of files at a time in rapid succession within minutes," and that Coral Sleet has run a workflow to trojanize private GitHub repositories since April. Microsoft is careful on one point, and so should we be: "AI-enabled operations remain human-directed and have not become fully autonomous." Microsoft Security blog

China's draft export controls for frontier AI treat agents as their own category, and comments close November 15. MOFCOM and the Ministry of Science and Technology issued Announcement No. 48 on October 16; a Cloud Security Alliance research note on it reached the Hacker News front page this morning. Per CSA's reading, model weights and related technology are restricted if a model hits any one of three tests (at least 10^26 training FLOPs, at least 2 trillion parameters, or a global top-10 ranking at export), while agent planning, multi-agent coordination and tool use are restricted above set benchmarks. Weights that are already public are outside the controls, but open-sourcing a new restricted model is not exempt. The catch for anyone building on Chinese open-weight models: CSA advises treating future release cadence as "a variable outside their control." CSA research note · MOFCOM announcement (Chinese)

New tools and features worth actually trying

The __usage_ai-training robots.txt group. Two lines today record a training preference that Google says it will read in early 2027 and that OpenAI, per secondary coverage, plans to read sooner. Add it under your existing per-bot groups, not in place of them. Honest tradeoff: nothing honors it yet on a date you can test against, and a more specific group for a named bot can still say something different.

dsh-superagent in observe mode. Worth reading even if you never deploy it, because the authorization doc is a clean template for "who may ask an agent for work." Its SETUP.md lists Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+ and a DeepSeek Harness install with a model provider. Honest tradeoff: workers run as your OS user without approval prompts, so it belongs on a dedicated machine or account.

stack-diagram v0.3.0. An agent skill that turns a plain description of your stack into an architecture diagram with real logos, exported as HTML, PNG, React or a Claude artifact. Honest tradeoff: it needs Python 3.9+ and network access to fetch logos, and PNG export pulls in Playwright with Chromium.

NOIR v0.2.5. A desktop tool that uses adb to audit an Android phone's apps, permissions and network endpoints, then scores the risk. Honest tradeoff: it needs USB debugging on, its findings are static evidence that "cannot prove a specific transmission occurred," and noir setup downloads adb and jadx even though the README says it never connects to the internet.

Trending AI repos on GitHub today

Read from Trendshift at about 07:20 ET; its ranks are momentum scores, not star totals. Stars below are cache-busted shields.io values.

  • deepseek-ai/dsh-superagent-prototype (#1): a research-prototype coding agent for one repo, built on DeepSeek Harness. Why now: one commit on October 31 and Bloomberg coverage on November 1. MIT, about 11k stars (the rendered GitHub page showed a stale 3.3k), no releases; "not an official DeepSeek product."
  • storytold/craft-anything (#2): a Rust engine and desktop app that studies a site, app or game and has a coding agent rebuild it. Why now: v0.2.0 on November 1. Apache-2.0 or MIT, about 2.4k stars; early alpha, and it needs a signed-in coding agent plus a fal.ai key for generated art.
  • modelflare-ai/AnyModelBot (#6): a local-first desktop agent whose models come through the operator's paid Modelflare gateway. Why now: v0.1.16 on November 2. MIT, about 1.2k stars; the README says data stays on your device, then says the gateway keeps request logs.
  • Verlintas/NOIR (#8): an adb-based Android privacy audit with risk scores. Why now: v0.2.5 on November 2. MIT, about 870 stars.
  • thedotmack/claude-mem (#10): a memory layer for Claude Code that records tool-use observations and feeds summaries into later sessions. Why now: v15.8.3 on November 2. Apache-2.0, about 110k stars; the "~10x token savings" claim comes with no method.
  • caru-ini/stack-diagram (#12): stack description in, logo-rich architecture diagram out. Why now: v0.3.0 on November 2. MIT, about 870 stars.
  • Thysrael/Horizon (#15): collects news, scores and dedupes it with AI, and writes daily briefings in English and Chinese. Why now: steady climb. MIT, about 12k stars, no releases; it shows a "Daily" badge while its own note says the daily workflow is disabled in the repo.
  • ccch1mneyyy/dsh-Go (#21): an unofficial Android client that runs DeepSeek Harness on the phone or drives a remote one. Why now: v1.0.9 on November 2, riding the dsh-superagent wave. AGPL-3.0, about 5.4k stars; not affiliated with DeepSeek.

What actually matters from today's signal

The trend to track is the move from settings to written rules. A training preference becomes a robots.txt group, an agent's permission to act becomes an allowlist and a sentence in a prompt, and a government's view of risk becomes a list of thresholds. For builders, the highest-signal areas this week are: your robots.txt, which now has a third kind of line in it besides "crawl" and "don't"; the authorization layer of any agent you let near a repo, where DeepSeek's observe-first default is worth copying; secrets held by third-party job runners, after Trigger.dev; and, if you depend on Chinese open weights, the November 15 comment deadline.

The counter-signal is that every one of these rules is voluntary for the reader. The training token works only for providers who implement it, on their schedule. DeepSeek's agent tells you in its own docs that its sandbox is not a boundary. And the attackers in Microsoft's report are running an open-source agent framework that reads no robots.txt and checks no allowlist. Write the rules, because the honest machines will follow them. Then put your real controls where they do not depend on good manners: rotated keys, branch protection, a separate OS account.

Source access notes: Vendor scan, Hacker News (Algolia, last 36 hours, 40+ points) and Trendshift read between about 07:15 and 07:25 ET. openai.com (news, the training-preference post and the misalignment-detection post) returned client errors, so OpenAI's position on the token is sourced to secondary coverage and marked as such; OpenAI's October 30 misalignment-monitoring post was left out because no primary text was reachable. The Verge, Wired and Bloomberg were blocked or disallowed; the DeepSeek Bloomberg item is cited only as coverage existing, not for any claim. Google's Search Central blog post returned headers only, so Google's position comes from its crawler documentation and Search Engine Roundtable's quotes of John Mueller. AI Weekly reported a Microsoft date (February 1, 2027) and an Anthropic implementation that Unite.AI's coverage did not support; both were left out. Anthropic's newsroom had nothing newer than October 28. Gemini 4 Carbon (October 28) and Mistral's Le Chonk (October 29) were covered in earlier briefings and skipped. The rendered GitHub page for dsh-superagent showed a stale 3.3k stars against 11k on cache-busted shields.io. Adversarial pass (Sonnet 5.5) caught four issues, all fixed: the headline and opening overstated OpenAI's commitment (only Google's is primary-sourced), a one-line versus three-line mismatch on the token group (it is two lines), a Trigger.dev quote spliced from two sentences, and dsh-superagent setup requirements that now cite SETUP.md.