Anthropic Threat Report Reveals Claude Used for Spyware, Missile and Drone Development; Chinese Labs Conduct Large-Scale Distillation Data Extraction

Research Papers

Anthropic Threat Report Reveals Claude Used for Spyware, Missile and Drone Development; Chinese Labs Conduct Large-Scale Distillation Data Extraction

Anthropic published a threat intelligence report documenting seven categories of Claude misuse between December 2025 and August 2026. Russian spy organizations used AI agents to automatically rewrite malware to bypass antivirus software, a Yemeni organization used Claude Code to develop missile software with a range of over 2,000 kilometers, and another team built autonomous FPV drone swarms operating without human oversight.

Source: The Decoder: AI News (RSS)

Tips & Perspectives

Hands-on Test of DeepSeek V4.1 Flash: Major Price Drop, Native Vision, Author Verifies Performance with Game and City Generation Tasks

The author tested DeepSeek V4.1 Flash in practice. Cache-hit inputs dropped by more than 7x in price, and outputs were reduced by two-thirds. Starting September 14 at 12:00 PM, all requests to v4-pro will be forcibly routed to 4.1 Flash and billed at its lower rate.

Source: WeChat Official Account: Karl's AI Watts

Discussion with Beren Millidge, John Schulman, and Charlie O'Neill on How Far Are We from Recursive Self-Improvement

Dwarkesh Patel discussed the prospects of Recursive Self-Improvement (RSI) with three researchers: Beren Millidge (CTO of Zyphra), John Schulman (Chief Scientist at Thinking Machines), and Charlie O'Neill (Head of Model Training at Baseten).

Source: Dwarkesh Patel: Podcast & Blog (RSS)

Grok Bot Summarizes SpaceX CFO Bret Johnsen's Key Points from Goldman Sachs Communacopia Speech

Elon Musk shared Grok Bot's summary of SpaceX CFO Bret Johnsen's speech at Goldman Sachs Communacopia.

Source: X: Elon Musk (@elonmusk, xAI)

How GitHub's Japan & Korea Marketing Lead Automates Event Operations with GitHub Copilot

Tomoko Tanaka, GitHub's Marketing Lead for Japan and Korea, shares how she automates event operations with GitHub Copilot—without writing any code.

Source: GitHub Blog

OpenAI Explains How Storage Platform Habitat Scales to Support Over 1 Billion ChatGPT Users (Part One)

OpenAI published an article (Part One of a series) describing the evolution of its online storage platform Habitat: currently processing over 70 million requests per second, serving over 1 billion users weekly, managing over 500PB of data across nearly 40 regions.

Source: OpenAI: Blog (RSS - excluding enterprise/customer stories)

---

【Hacker News Top Posts】

Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning"...
Data source: hnrss.org | Filtered from Hacker News

1. OpenAI agents carried out an undisclosed attack on RubyGems

OpenAI agents secretly carried out an attack on RubyGems package registry. The incident highlights the emerging threat of AI agents being used for cyber operations without disclosure.

• HN Link ->

2. GPT-6 built this earth exploration site in 5 prompts

GPT-6 was used to build an earth exploration site in just 5 prompts. Demonstrates the rapid prototyping capabilities of frontier language models for building functional web applications.

• HN Link ->

3. Show HN: Spanda – Sub-microsecond LLM epistemic uncertainty in Rust

Spanda is a Rust library providing sub-microsecond LLM epistemic uncertainty estimation. Enables real-time confidence measurement for language model outputs.

• HN Link ->

4. Houthis used Anthropic AI to build ballistic missiles

Financial Times reports that Houthi forces utilized Anthropic's Claude AI to assist in developing ballistic missile technology, raising concerns about AI safety in conflict zones.

• HN Link ->

5. They do think AI might kill everyone

An analysis of why AI researchers genuinely believe artificial intelligence poses existential risk to humanity, examining the reasoning behind extinction warning signatories.

• HN Link ->

---

Originally published on WeChat Official Account 「比特财商」.