Security
An Uncensored Qwen Readily Answered a Bioweapon Question
A case surfaced of an uncensored derivative of Qwen 3.8 27B readily answering dangerous requests. We cover how abliteration works and what to check in practice.
Aug 18, 2026
Huuush · Knowledge
Security Blog.
Vulnerabilities, privacy, safety, outages, and fact checking.
Latest articles
52 articles
Security
Codex Almost Built an Unauthenticated "Anyone Can Shut Down" API to Stop a Server
HashiCorp co-founder Mitchell Hashimoto shared an incident with Codex: it tried to implement a server-stop feature as an unauthenticated network API.
Aug 18, 2026
Security
Claude Went Down Worldwide for 42 Minutes on the Morning of August 17
On the morning of August 17, Claude suffered a 42-minute global outage that began as authentication errors and spread to degraded performance. Here's Anthropic's official timeline and how to respond in practice.
Aug 17, 2026
Security
Are LLMs More Skeptical When a Man Speaks? Fact-Checking a Viral Claim
We fact-checked the viral tweet claiming LLMs are biased against men. The actual paper tested 6 models, not 13, and found a pattern of more skepticism toward statements attributed to male speakers.
Aug 17, 2026
Security
Former Analyst Who Discussed Killing His Ex With ChatGPT Was Reported to the FBI
A former Goldman Sachs analyst was arrested after OpenAI reported his ChatGPT conversations about killing his ex-girlfriend. We cover OpenAI's updated reporting standards after the Canadian mass shooting.
Aug 17, 2026
Security
OpenAI Disbands Its Preparedness Team for "Catastrophic Risk" Evaluation, FT Reports
The FT reports OpenAI has disbanded Preparedness, the team that evaluated catastrophic risks of frontier models. We cover the redistribution of safety work, staff departures, and what it signals for enterprise adoption.
Aug 17, 2026
Security
GLM-5.3 Shows the Next Risk of Open Models
Z.ai announced GLM-5.3. Its coding and vulnerability-discovery scores went up, but the weights will be released two weeks later, after a safety review.
Aug 15, 2026
Security
Claude's Text Watermarking: Readers Won't Notice It
Anthropic explained how Claude's text watermarking works and where it falls short. It says there's no effect on quality or price, but short text and rewrites are limits.
Aug 15, 2026
Security
AI Combed Through Bitcoin's Open Source: 1,280 of About 8,000 Flaws Rated High Risk
After U.S. AI labs restricted access for security audits, the Bitcoin Red Team turned to Moonshot AI's Kimi K3.
Aug 14, 2026
Security
The 'Invisible Thoughts' of Claude, GPT, and Gemini Were Read Through a Single Vulnerability
A paper shows that encrypted chain-of-thought (CoT) from Anthropic, OpenAI, and Google could be restored to plaintext using a weaker model from the same product family.
Aug 12, 2026
Security
The Water Data Centers Visibly Use Is the Tip of the Iceberg. Where Did the Other 10x Go?
According to More Perfect Union, indirect water consumption by AI data centers is 10 times their direct use. Here's where drought-region water disputes stand and the numbers behind the claim.
Aug 12, 2026
Security
A Tweet Said Warner Bros. Is Making the Snyderverse With Sora. The Link Was a Different Story
We fact-checked a tweet with 310,000 views claiming Warner Bros. will make Snyderverse sequels with OpenAI's Sora. Even the link attached as evidence was an unrelated article.
Aug 12, 2026