One robots.txt Line for AI Training: OpenAI, Microsoft, Google and Anthropic Back a Shared Signal
On October 30, four AI providers said they will honour a common robots.txt token that tells every participating crawler a site does not want its content used to train generative models. The syntax is in an IETF draft. Microsoft adds a separate grounding control and fixed dates for when it starts reading the new signals.
A Tampered Claude Code Extension Sat in the Cursor Marketplace for 11 Hours
Cursor says a modified copy of Anthropic's Claude Code extension, which Anthropic did not build, was available in its VS Code extension marketplace on October 29 and tried to install a persistent SSH backdoor. Coding errors stopped the backdoor on most machines, but the persistence and the reconnaissance still ran.
Google's Suncatcher Satellite Ran Gemma 4 in Orbit, One TPU at a Time
TechCrunch reported on October 30 that Google's Project Suncatcher prototype, launched in early October with four TPUs, has finished commissioning and answered its first prompts. Power limits mean one chip can run for about 20 minutes at a time, and the real test is how radiation affects the hardware.
New York Bans Six Chatbot Behaviours for Minors as OpenAI Tightens Its Teen Limits
Governor Hochul signed S.9051B/A.10379C on October 28, barring chatbot features such as pretending to be human or encouraging secrecy for anyone not confirmed as an adult. Two days later OpenAI added weapons limits for teens and an automatic end to chats that fixate on violence. Both rely on knowing who is a minor.
Framework Watch: Pydantic AI Stops Sending Server State to the Wrong Endpoint
Pydantic AI 2.64 scopes stored response and container IDs to the endpoint that created them and adds a scanner for planted instructions in memory and skills. ADK Python stops storing the model's hidden reasoning as memory. LangGraph 1.2.19 fixes trace settings that emptied stream output, and Gemini CLI fixes symlinked paths in its glob tool.
Research Radar: Making a Model Commit to a Safety Decision Before It Answers
Preprints submitted October 29 turn shallow safety alignment into a defence against multi-turn attacks, treat checking a computer-use agent's work as a task for another agent, rescue reinforcement learning when every attempt fails, let a weak verifier rewind a strong prover, and find that data attribution improves when it stops averaging.
Trend Watch: A Name Is Not an Identity, and This Week Proved It Four Times
From October 29 to 31, a tampered extension shipped under a trusted publisher's name, a framework sent stored state to any endpoint sharing a provider name, a glob tool returned paths that named the same file differently, and a memory service treated hidden reasoning as conversation. The fixes all check what something is, not what it is called.