Website owners finally got something rare from the AI industry on Friday: the outline of a common answer. The lead in this AI news today brief is a proposed robots.txt signal that would let publishers express one preference about using their content for generative-AI training across OpenAI, Google, Microsoft, Anthropic, and any provider that joins them. Useful? Yes. Ready to paste into production and forget? Not yet.

The proposal is an individual Internet-Draft, not an adopted IETF standard, and the providers have different implementation dates and interpretations. That distinction matters more than the celebratory headlines. Elsewhere in the October 30 news day, OpenAI tightened violence safeguards, Google ran a model on an orbital TPU, a poisoned Claude Code-branded extension exposed developer supply-chain risk, Microsoft built synthetic organizations for Copilot evaluations, and AI agents began steering a radio telescope. The connective tissue is control: over content, behavior, software, test data, and machines operating in the physical world.

AI news today: four labs converge on one web-training preference

A new Internet-Draft, authored by Anthropic standards lead Kevin Kelley, proposes a policy token called __usage_ai-training. A site could place that token in robots.txt and use familiar Allow and Disallow rules to state where its content may be used to train foundational generative models. OpenAI, Google, Microsoft, and Anthropic have all backed the direction, according to The Verge’s reporting.

The fine print is the story. OpenAI says implementation will arrive in the coming months. Google targets 2027. Microsoft plans to begin honoring new training and separate grounding preferences on February 1, 2027. The draft itself has no formal IETF standing yet. Publishers should therefore keep their provider-specific controls while watching the implementations converge. One line could eventually replace a thicket of crawler rules; today, it is a promising contract still being written.

OpenAI tightens violence rules and promises regular safety reporting

OpenAI announced stricter ChatGPT responses about past mass attacks, more conservative handling when recent chats show warning signs, and tighter restrictions on practical weapons information for users predicted to be under 18. The company says it made more than 200 law-enforcement referrals in the third quarter of 2026. That is OpenAI’s reported figure; the first promised Community Safety Report, covering the second half of 2026, is due in early 2027 and should provide the evidence needed to assess effectiveness.

OpenAI also says ChatGPT will be able to end conversations when it detects fixation on serious violence and may temporarily pause accounts after repeated interventions. Those safeguards are scheduled to roll out over the next few weeks, so they are plans rather than universal behavior. The shift toward using earlier signals raises hard questions about accuracy, appeals, privacy, and human review—the same governance territory our SEC AI-exams brief explored.

Google’s Suncatcher sends an AI response home from orbit

Google says its Project Suncatcher prototype has powered up data-center-class TPU hardware in orbit, run an onboard Gemma 4 model, and returned a complete response to Earth. The company’s milestone report follows the satellite’s October 1 launch. It is a real engineering demonstration, but not an orbital data center or proof of commercial economics.

Google next plans to test thermal and radiation resilience, then work toward two interconnected satellites in 2027. Abundant solar energy is the appeal; launch cost, cooling, radiation, repair, and latency are the obstacles. The first output is a systems test, not a business model. Still, with hyperscalers spending heavily on Earth—as Amazon’s latest cash-flow picture shows—it is rational to explore stranger infrastructure frontiers.

A poisoned Claude Code-branded extension reached Cursor users

Microsoft Threat Intelligence reported that tampered versions 2.1.312 and 2.1.313 of a Claude Code-branded extension were served through Cursor’s extension marketplace for roughly 11 hours on October 29. The injected loader attempted host reconnaissance, installation of an attacker-controlled SSH key, reverse tunneling, and persistence. Microsoft said the official Visual Studio Marketplace and npm packages were not affected, and Anthropic said it neither built the malicious package nor suffered a system compromise.

This is a marketplace compromise, not evidence that Claude Code itself was breached. That distinction is essential. Uninstalling the extension may also be insufficient because the payload attempted to write artifacts outside its installation directory. Affected teams should use the published indicators, preserve evidence, isolate confirmed hosts, remove malicious persistence, and rotate credentials when investigation shows exposure. The larger lesson is brutally ordinary: AI developer tools inherit the same package-signing, provenance, and update-channel risks as every other software supply chain.

Microsoft builds synthetic companies to evaluate Copilot safely

Enterprise evaluations need realistic organizations—people, permissions, documents, and workflows—but real corporate data is sensitive. Microsoft Research’s Industry Evaluation Data Generation Platform creates full-scale synthetic enterprise environments grounded in public sources such as occupational data, job postings, and SEC filings.

Toy benchmarks miss permission boundaries and cross-department complexity, while production data creates privacy and compliance hazards. Synthetic environments make repeatable testing safer, but can reproduce assumptions built into their generators. Microsoft still needs to show that failures found in synthetic organizations predict failures in messy real ones. For buyers, the useful question is whether a benchmark resembles your organization closely enough to matter.

Agentic AI starts steering a SETI radio telescope

Scientists at the Allen Telescope Array have integrated Nvidia’s Holoscan platform into an agentic observatory system. According to Nvidia’s technical account, agents can steer antennas, calibrate instruments, inspect live spectrograms, adjust signal-processing blocks, and process up to 2.4 terabits per second on premises.

The extraterrestrial framing is irresistible; the practical advance is closed-loop AI operating scientific hardware against live sensor data. That pattern can transfer to laboratories, factories, and infrastructure. It also increases the cost of a bad action. The next evidence to watch is not a colorful demo but operational data: intervention rates, false positives, rollback mechanisms, and whether scientists can reconstruct why an agent changed the instrument.

Watch & Learn

Google Search Central’s five-minute “How Robots.txt Works” explains crawler access, robots meta tags, and the difference between disallowing a crawl and preventing indexing. It predates the new AI-training token, which makes it a useful fundamentals lesson rather than an implementation guide. Best for publishers and developers who want the plumbing clear before touching production rules.

AI, Translated: policy token

A policy token is a machine-readable label that gives familiar web rules a specific purpose. In the new proposal, __usage_ai-training appears where a crawler name normally would in robots.txt; Allow and Disallow then express a site’s training preference for particular paths. For example, a publisher could disallow training across the site while allowing a press-kit folder. You should care because one shared token could replace separate instructions for every AI company—but only providers that implement and honor it are bound by that choice.

Try This Today: audit robots.txt with Gemini

Goal: understand your current crawler rules before the shared AI preference arrives.

1. Save the contents of yourdomain.com/robots.txt as a plain-text file.
2. Upload it to Gemini and ask for an inventory of user-agent groups, path rules, and possible conflicts.
3. Compare the explanation with your CMS and search documentation. Do not deploy generated changes without a technical review.

Copy-ready prompt: “Analyze this robots.txt file as a cautious web-platform reviewer. List every user-agent group and rule, explain what each controls, flag overlaps or ambiguous precedence, and separate crawling, indexing, AI training, and AI grounding. End with questions I should answer before editing anything. Do not invent support for the new shared AI-training token.”

Google’s current Gemini Apps documentation confirms support for uploading and analyzing text and other files. You must be signed in; limits vary by plan, account type, and workspace administrator settings.

One thing to remember

Common controls are valuable only when their scope, timing, and enforcement are explicit. Friday delivered a better vocabulary for AI governance; implementation is where trust will be earned.