🤖 AI HOT Daily · Oct 9, 2026

Today's focus: CrowdStrike reports that a likely single Chinese-speaking actor used the AI-driven open-source penetration tool ARTEX from late September into early October to breach multiple South Korean financial institutions and steal large amounts of data, including over 25,000 Shinhan Bank records with names, contact details, income and credit limits. Industry and revenue: the Financial Times reports OpenAI told investors annualized revenue neared $50B through September, well below outsider estimates of $70B — a roughly $20B gap driven by accounting scope, since Anthropic counts partner revenue from AWS and Google Cloud while OpenAI excludes it — sending US tech stocks down; Codex and ChatGPT Work combined hit a record 40 million active users with paid-account resets fully credited; Waymo closed a $5B term loan, its first debt financing (PIMCO, Blackstone and Sixth Street leading, Goldman Sachs sole lead bookrunner), to accelerate its robotaxi service in the US and abroad; and OpenAI banned two covert influence operations — Russia-linked Dark Clark, whose fake persona Mia Clark ran the Latin American think tank Social Research Center and earned a first-ever Category 5 rating, and Iran-linked Bogus Bylines, which used seven fake journalists. Models and products: Anthropic launched Claude Dashboards and Claude Motion in beta; released OSS Scanner, an opt-in service using its strongest models (including Claude Mythos) to periodically and freely scan open-source projects with fully model-generated, human-unreviewed reports; published its 2026 usage policy update effective Nov 12 with a new deceptive-activity section, a narrowed election clause, explicit bans on weapons software and arming drones, sharper monitoring limits, human-in-the-loop requirements for autonomous physical operations, and a ban on persistent gratuitous abuse of models; and ts-rust ports microsoft/TypeScript's (Go) compiler, type checker and language server to Rust, with the author saying an LLM wrote all the code and he hasn't read it. Research and opinion: the mathematician group AHM, under Terence Tao, issued a statement calling for a boycott of OpenAI after it released 700-plus AI-generated proof files at once, calling it a display of force rather than scholarship (OpenAI had claimed its internal model solved 100+ open problems in a month with ~5% success on ~8,000 test problems); Goodfire built and deployed a production cyber-security monitor cascade of activation probes plus an LLM judge on Kimi K3 and GLM 5.3; and a Hugging Face engineer used ML Intern to build seven small models in days for about $103, including a CPU-runnable 0.8B prompt rewriter with 99.7% valid output and a Qwen3.5-2B that lifted citrus-disease recognition from 14.9% to 52.8%. Also: Arena posted Claude Haiku 5.5 (High) debuting 30th in Code Arena at 1587; Zenity found a single prompt hijacks every AgentCore agent in an AWS account; Artificial Analysis ranked Google Nano Banana 2.1 fourth on two leaderboards at half its predecessor's price; Google open-sourced the ML Drift edge GPU inference engine replacing the TFLite GPU delegate; Arena raised a $200M Series B at a $3.1B valuation and launched an Alignment Index; GPT-6.1 Sol ultrafast shipped; Anthropic published a guide to scheduled agent automation with Claude Managed Agents; Tessl wrote on three key design decisions for enterprise agent memory; Google open-sourced the AQuA ambient quality agent; and Claude's September recap merged Chat and Cowork and launched the Claude 5.5 series.

1. 1. CrowdStrike: AI-Powered Tool Breached Multiple South Korean Banks

   A likely solo actor used the open-source ARTEX tool from late September; 25,000+ Shinhan Bank records with names, income and credit limits leaked.

2. 2. Claude Adds Dashboards and Motion

   Anthropic released real-time Dashboard and animated-explainer Motion features in beta.

3. 3. Anthropic Launches the OSS Scanner for Open Source

   Its strongest models including Claude Mythos periodically scan open-source projects for free, producing fully model-generated, unreviewed reports.

4. 4. ts-rust: An LLM Ported the TypeScript Compiler to Rust

   Compiler, type checker and LSP moved to Rust; the author says an LLM wrote it all and he hasn't read the code.

5. 5. OpenAI Annualized Revenue Nears $50B, ~$20B Below Expectations

   The FT ties the gap to accounting scope — Anthropic counts AWS/Google Cloud partner revenue, OpenAI excludes it — sending US tech stocks lower.

6. 6. Mathematicians Call for an OpenAI Boycott as AI Proofs Flood In

   AHM, chaired by Terence Tao, calls the one-shot release of 700-plus AI-generated proof files a display of force rather than scholarship.

7. 7. Anthropic's 2026 Usage Policy Update Takes Effect Nov 12

   Adds a deceptive-activity section, narrows elections, bans weapons software and drone arming, requires human-in-the-loop for autonomous physical tasks, and bars persistent model abuse.

8. 8. Codex and ChatGPT Work Hit a Record 40M Active Users

   Paid-account resets are fully credited, with GPT-6 live in Chat.

9. 9. OpenAI Bans Two Russian and Iranian Influence-Operation Clusters

   Russia's Dark Clark ran a LatAm think tank via a fake persona for a first-ever Category 5; Iran's Bogus Bylines used seven fake journalists.

10. 10. Waymo Closes $5B Debt Financing

    Its first debt raise, led by PIMCO, Blackstone and Sixth Street with Goldman as sole lead bookrunner, to accelerate robotaxi expansion.

11. 11. Goodfire Deploys Cyber Monitors on Kimi K3/GLM 5.3; HF Builds 7 Models for $103

    Goodfire cascades activation probes with an LLM judge; HF's ML Intern produced a CPU-runnable 0.8B rewriter (99.7% valid) and a Qwen3.5-2B lifting citrus-disease accuracy from 14.9% to 52.8%.

12. 12. More AI News: Zenity's One-Prompt AWS Agent Hijack, Arena at $3.1B, and More

    Haiku 5.5 debuts 30th in Code Arena at 1587; Zenity hijacks all AgentCore agents with one prompt; Nano Banana 2.1 ranks 4th at half price; Google open-sources ML Drift and AQuA; Arena raises $200M Series B; GPT-6.1 Sol ultrafast ships; Claude Managed Agents guide; Tessl on agent memory; Claude's September recap.