Tuesday, November 3, 2026

TL;DR

Anthropic's amended IPO filing admits future models may imminently cross the threshold where AI speeds up its own research, and that it might not notice until afterward. Investors are pushing its valuation down to $1.5-1.6 trillion, from $2 trillion.

Act on this

  • If your infrastructure logs OPENAI_API_KEY, audit for exfiltration: Cisco Talos found live NIK C2 malware harvesting that key to run AI analysis on stolen data. Cisco Talos, NIK C2 #

Signals

A frontier model can now edit another model's weights directly from plain-language goals, skipping gradient descent - though named researchers dispute how much that actually adds.  #

MIT CSAIL's Rikiya Takehi and coauthors showed Nov 2 that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and raised its MATH-500 score from 7.9% to 48.1% - but call the result 'very early' and say Astra added no ability Gemma 3 270M lacked. Interpretability researcher Adam Karvonen called it ablation of a loop, not a weight rewrite, per Techmeme's aggregation of The Register and VentureBeat. Rikiya Takehi, Sparks of Weight-Space Programming · swyx/Latent Space, AI News on Weight-Space Programming

AI-integrated malware is spreading beyond isolated proof-of-concept builds into a wider toolkit of live criminal frameworks.  #

Cisco Talos found two new AI-enabled threats active in the wild this week: NIK C2 (Oct 29), whose implant calls the OpenAI API directly from a victim's machine to analyze stolen data, and VOLTA (Oct 31), an infostealer that uses an LLM to rank which victims' assets are most valuable - a different build from September's CLOSEDQUORUM, not a repeat of it. Cisco Talos, NIK C2 · Cisco Talos, Appraisal by AI (VOLTA)

News

WATCHAnthropic says it may be nearing the threshold where AI speeds up its own research  #

Its Nov 2 amended S-1 says internal forecasts for future models, not current ones, indicate Anthropic may 'imminently' meet its Responsible Scaling Policy's bar for AI-accelerated progress - and that it might not detect the shift until after it happened. TechCrunch, Anthropic warns that it's very close to RSI capability

WATCHInvestors push Anthropic's IPO valuation down to $1.5-1.6 trillion, from $2 trillion  #

Q3 revenue hit about $17 billion, but growth slowed to 47% sequential from 144% in Q2, per the amended S-1. The Wall Street Journal reports institutional buyers now want $1.5-1.6 trillion, citing a broader tech selloff and doubts about AI returns. Dataconomy, Investors cut Anthropic IPO valuation target to $1.5 trillion

SHIPCrowdStrike ships an AI Kill Switch to cut off rogue agents' model access  #

Security teams can now order a connected AI gateway to block the models powering a misbehaving agent's decisions, via Falcon Guardian. CrowdStrike also announced plans to acquire Fiddler for AI agent observability and Dreadnode for automated AI security testing. CrowdStrike, Introducing AI Kill Switch

SHIPGoogle DeepMind opens a research preview of Genie 4, a real-time world model with native audio #

Genie 4 generates interactive 3D environments from a prompt, holds a consistent scene for up to two minutes - double its predecessor's limit - and produces matching ambient sound and dialogue. Access is limited to trusted testers; Google names no launch date or price. Northeast Times, Google DeepMind opens Genie 4 preview to testers

WATCHRealtime-Bench finds a large gap between what agents can solve and what they can solve in time #

UT Austin, Ohio State, UIUC, Nvidia and Northwestern researchers' new benchmark found top models solve 70-80% of dynamic tasks with the clock paused, but only 36% in real time, per Jack Clark's Import AI 479. Jack Clark, Import AI 479

WATCHThe Census Bureau quietly renamed its AI survey question to 'Super Intelligence' #

Simon Willison found the Business Trends and Outlook Survey's Oct 19 wording change, tied to Trump's September executive order, makes its widely cited AI-adoption trend data incomparable across the break. Simon Willison, The U.S. Census Bureau and 'Super Intelligence'

The long view

A year ago, Anthropic's Responsible Scaling Policy already named an AI R&D acceleration threshold as a future risk - a line in a safety document, not something any lab said it was close to. In its Nov 2 amended IPO filing, Anthropic said its own forecasts for future models now indicate it may imminently cross that threshold, and that it might not detect the acceleration until after it happened. If this holds, the real test is whether Anthropic names the threshold crossed in a system card in real time, not after the fact in a future filing.

Also noted