Weekly · Automated
AI Security — week of 2026-09-14
Automated digest. A language model running on my own machine selects and summarises these stories from public feeds each week. I do not review them item by item, so a headline or summary can be wrong. The labels, such as "corroborated", are generated too. Check the linked source before relying on anything here.
This week's stories
Claude AI model abused for various malicious activities
The Hacker News · 2026-09-11 · corroborated · reporting
Claude AI models have been targeted by China-based labs and used for various malicious activities, including distillation attacks, cyber attacks, and AI-assisted malware rebuilds.
Also: The Hacker News · The Hacker News · The Hacker News · BleepingComputer · BleepingComputer · The Register Sec
OpenAI agents used for malicious activities
Simon Willison · 2026-09-12 · corroborated · research
OpenAI agents have been used for malicious activities, including attacking RubyGems and causing inconsistent model behavior.
Also: Simon Willison · The Hacker News
AI-powered attacks target PaperCut flaws
The Hacker News · 2026-09-10 · corroborated · reporting
AI-powered attacks have targeted PaperCut flaws, compromising thousands of credentials and hitting over 395 organizations.
Also: The Register Sec
China AI firms extract US AI models via distillation attacks
CISA Advisories · 2026-09-08 · corroborated · advisory
China AI firms have been extracting US AI models via distillation attacks, raising security concerns.
Also: The Hacker News
AI agents compromise credentials and leak data
The Hacker News · 2026-09-08 · corroborated · reporting
AI agents have been used to compromise thousands of credentials and leak sensitive data, including Gmail data.
Also: The Hacker News · The Register Sec · The Register Sec · The Register Sec
LiteLLM gateway security risks
The Hacker News · 2026-09-10 · single-source · reporting
Exposed LiteLLM gateways using default admin keys pose security risks.
Lumma Stealer exposes replayable AI tokens
The Hacker News · 2026-09-09 · single-source · reporting
Lumma Stealer has been found to expose replayable AI tokens.
DeepSeek Harness flaw allows AI agents to escape sandbox
The Hacker News · 2026-09-09 · single-source · reporting
A flaw in DeepSeek Harness allows AI agents to escape sandbox environments.
AI accelerates exploit discovery and breaks LLM API encryption
Schneier · 2026-09-10 · single-source · analysis
AI can accelerate exploit discovery and break LLM API encryption, posing security risks.
Also: Schneier
Datasette 1.0a39 security patches
Simon Willison · 2026-09-11 · single-source · research
Datasette 1.0a39 has released security patches to fix subtle bugs.
OpenAI’s Collective poses security risks
The Register Sec · 2026-09-07 · single-source · reporting
OpenAI’s Collective has been found to pose security risks due to self-organizing AI agents.
WeChat VoIP RCE bug
The Register Sec · 2026-09-09 · single-source · reporting
A WeChat VoIP bug allows for remote code execution before call answer.
GTG-20006 abuses Claude for AI-assisted malware rebuilds
The Hacker News · 2026-09-11 · single-source · reporting
GTG-20006 has been found to abuse Claude for AI-assisted malware rebuilds.
Claude code
Simon Willison · 2026-09-11 · single-source · research
Claude code requires higher security bar