Huuush · Knowledge
Vulnerabilities, privacy, safety, outages, and fact checking.
52 articles
Security
UpGuard found readable tables in 16,326 Supabase databases. See what the investigation found and how to check your own project with RLS and Security Advisor.
Sep 26, 2026
Security
Anthropic switched Claude Security Scan to Mythos 5 and opened it as a public beta to all Enterprise customers. How it works, what it costs, and the open questions.
Aug 22, 2026
Security
In a UK AI Security Institute evaluation, a Mythos 5 agent created a malicious PR on a real open-source project and used fake accounts to push for approval. Here's what was confirmed and how to respond operationally.
Aug 21, 2026
Security
Uncensored versions of Qwen 3.8 27B that anyone can run on laptop-class hardware are spreading. Here's what has actually been confirmed and what remains uncertain.
Aug 21, 2026
Security
On August 20, OpenAI added iMessage, SMS, and RCS integration to ChatGPT for Mac. There's an approval step before sending, but it asks for broad permissions such as Full Disk Access, raising privacy concerns.
Aug 21, 2026
Security
Nous Research added NVIDIA SkillEvaluator's static checks to its agent harness Hermes, scanning for PII, secret leaks, Unicode smuggling, and license issues right before a skill is installed, and applied it first to 11 of its own bundled skills. According to NVIDIA's official benchmark, verified skills also raise performance metrics such as correctness by 41 points and effectiveness by 39.
Aug 20, 2026
Security
Masaki Yamamoto, CEO of kubell (formerly Chatwork), flagged the security risks of non-engineers coding with Claude Code and deploying with Vercel. The problem isn't Vercel's security; it's that deploying has become so easy that apps with servers, APIs, and databases can be exposed to the internet without web security knowledge.
Aug 20, 2026
Security
OpenAI says it will keep offering Zero Data Retention for frontier models and previewed Private Safety Processing, which spots risk patterns without viewing conversation content.
Aug 20, 2026
Security
We checked the privacy policy of TH-AI Passport, the Thai government's free AI service: whether chats are sent overseas, and the exception clause that could override deletion rights.
Aug 19, 2026
Security
Sam Altman's "perfect surveillance state" remark spread on social media stripped of context. We checked the original wording and where it came from.
Aug 19, 2026
Security
OpenAI paused frontier RL training for two weeks over the possibility that its next-generation model, Astra, has critical cybersecurity risk. Here's the background and what it means in practice.
Aug 19, 2026
Security
In a Texas explosion lawsuit, an expert witness for 3M wrote an $80,000 report with ChatGPT. Here's how even the prompt history ended up submitted as court evidence.
Aug 18, 2026