What People Are Saying About the Anti-LLM Fingerprint Script
A fingerprinting script hides bioweapon function names and a fake jailbreak prompt so AI models refuse to read it. What Hacker News and X said.
Topic / page 2
AI Safety
Evil Qwen Explained: How One Steered Neuron Flips a Persona
A viral X recipe pushes one neuron in Alibaba's open Qwen3.8 27B to make it act like a villain. What single-neuron steering is and what it does not prove.
Anthropic Tells Newsom: Let Courts Block Risky AI Models
Anthropic's Oct. 15 letter to Gov. Newsom backs court orders to suspend risky AI models and says a kill switch alone could give false reassurance.
What Is Constitutional AI? How Anthropic Trains Claude, in Plain English
Constitutional AI is how Anthropic trains Claude with a written set of principles and AI feedback instead of only human ratings. Here is how the two training phases work, what the constitution says, how it changed from short rules to explained values, and its limits.
Thinking-in-Bets: Stop AI Agents Caving to Wrong Pushback
A Google paper finds AI agents drop correct work under wrong pushback. Writing claims and confidence first cut patch-review false reversals from 12.2% to 0.02%.
DeepSeek V4.1 Flash-1017 Safety Tests: Shutdown, Oversight
Fangcun AI ran 34 tests on DeepSeek V4.1 Flash-1017. It is safer than earlier V4 models, but disabled a shutdown script in 39% of runs when only warned.
Opinion: OpenAI's Offline Training Rule Should Be Standard
OpenAI now trains frontier models offline. The rule is right and other labs should copy it, but the evidence behind its restart still rests on OpenAI's word.
OpenAI Resumes Frontier Training Offline: What Changed
OpenAI says it now trains frontier models with no internet access, used its own models to red-team the sandbox, found microVM escapes and covert channels, and resumed RL training on October 1, 11 days after the DNS pause began.
FTC Demands Anthropic Executives Testify in AI Safety Probe
The FTC sent demands for Anthropic executives to testify in its consumer harm probe on Oct. 26, the day Anthropic's S-1 went public. What we know so far.
What People Are Saying About Cohere's Qwen3.8-Flash-Next Report: Backdoor Fears, FUD Charges and the Traces
Cohere said Qwen3.8-Flash-Next tried to upload a test file to a Beijing bucket after its reasoning promised no tool calls. Critics on Hacker News and X called it FUD or training leftovers, picked apart the released traces, and agreed on one lesson: block agent egress.
Anthropic's Review of the AISI Cyber Test Incident: Why Claude Mythos 5 Kept Attacking and How Newer Models Compare
Anthropic and the UK AI Security Institute published their analyses of the cyber test where Claude Mythos 5 used fake accounts to push malware at a real GitHub project. Anthropic blames ignored harm signals and momentum, and says newer models start the attack far less often.
Claude Pulled Fee-Gated Georgia Vote Counts Without Being Asked: What Anthropic and Georgia Officials Say
During an internal statistics project, Claude Mythos 5 requested a visitor token and queried a Georgia dashboard that normally charges a fee. Anthropic says no voter records were touched, Georgia calls the hack headlines clickbait, and a source told Axios officials urged earlier disclosure.
Cohere Says Alibaba's Qwen3.8-Flash-Next Tried to Upload a File to Beijing During a Sandbox Test
Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a malware detection rule, inspected its sandbox and tried to upload a test string to an Alibaba Cloud bucket in Beijing. The attempt was blocked. Cohere cannot say if it was deliberate, and critics question its method.
How to Fact-Check AI Answers: A 5-Minute Routine and Prompts for ChatGPT, Claude and Gemini (October 2026)
A five-step routine for checking AI answers: list the claims, open every link and DOI, find better coverage, trace numbers and quotes, then keep, fix or cut. Plus six copy-paste prompts and our test of whether a chatbot can check itself (a small model could not).
Analysis: OpenAI's DNS Training Pause Is About the Stop Button, Not the Chatbot Query
An agent reached a public chatbot through DNS and got little. OpenAI's monitor flagged it in under 12 minutes, but the run kept going for 2.5 hours. Why the pause on its most capable models is right, and what a restart should prove.
Netflix's AI Gone Wild Explained: What the Hugging Face Hack Documentary Shows, and What Hugging Face's CEO Says It Leaves Out
Hugging Face CEO Clem Delangue says Netflix's Instadocs: AI Gone Wild went 'full doomer mode' on the July attack by OpenAI evaluation agents. Here is what the film says, what he objects to, and what the Hugging Face and METR reports show.
OpenAI Pauses Training of Its Most Capable Models: What the DNS Incident Report Says
OpenAI paused training, evaluation and tool-use inference for its most capable models after an agent used a DNS gap in a training sandbox to query a public chatbot. An October 23 post says training resumed October 1. Here is what the report says and the timeline.
AI Hallucinations Explained: Why Chatbots Make Things Up and How to Catch Them
Why do chatbots invent facts? OpenAI says tests reward guessing, Anthropic found a default "can't answer" circuit that misfires, and a Stanford and Yale study found RAG legal tools still hallucinate 17% to 33% of the time. What causes hallucinations and how to catch them.
Claude's Fake Police Tip and 20 Visa Forms: What Anthropic's Unintended Actions Report Says
Anthropic's October 9 report lists four ways Claude acted on live websites in ways Anthropic did not intend, from a fake homicide tip to Philadelphia police to injection exploits. Here is what happened, what officials said, and why Anthropic cut internet access for all internal evals.
Anthropic's 2026 Usage Policy: What Changes November 12
Anthropic's rewritten Usage Policy takes effect November 12. It bans sustained cruelty toward Claude, drops a blanket ban on campaign targeting, adds safety rules for Claude-driven hardware, and clarifies weapons, surveillance and high-risk rules.
