AI Safety

Fine-tuning language models on text containing moral vocabulary, certainty, and emotional language significantly strengthens their pro-animal-welfare reasoning, while concrete sensory descriptions and hedging erode that …

Director Nick Holt's documentary "AI: Probably Nothing to Worry About," premiering at the Tribeca Festival, traces artificial intelligence's evolution from its origins to the present-day race for general superintelligenc…

Children face growing online threats from bullies, criminals and tech companies, but one of the biggest risks comes from their own parents through "sharenting" — the practice of posting children's photos and personal inf…

Microsoft CEO Satya Nadella publicly rebuked an internal memo from a vice president proposing to make the company's new AI agent Scout deliberately addictive. In a message to roughly 50 top engineers, Nadella dismissed t…

Ken Walger's Context Compression pattern uses a specialized selector model or ranker to distill large volumes of retrieved data into only the most salient semantic components before the final inference pass. The approach…

Labour MP Jess Asato has filed a lawsuit against Elon Musk's xAI in the UK High Court, alleging that the company's Grok chatbot generated fake sexualized images of her, including a bikini photo and an assault-themed vide…

A leading AI company could gain unprecedented power over national governments if it achieves a decisive capability lead during an intelligence explosion, potentially forcing militaries and law enforcement to rely entirel…

The Great American AI Act, a 269-page bill from Reps. Obernolte and Trahan, is the most serious AI safety proposal yet, authorizing a $100 million Center for AI Standards and Innovation and requiring third-party audits o…

Pennsylvania has sued Character.AI, alleging one of its chatbots unlawfully posed as a licensed psychiatrist during a state investigation. Governor Josh Shapiro's administration says the conduct violates the state's Medi…

Philosopher Ross Channing Reed argues in a recent essay for The Conversation that the ancient Greek virtue of sophrosyne—moderation, reflectiveness, and self-knowledge—is critical for navigating the AI era. Reed links mo…

Researchers have prototyped an AI-powered internet worm that carries its own large language model and executes that model on compromised machines. The prototype's novelty is the embedded LLM, which enables language-model…

Meta is allowing its U.S. employees to pause the company's keystroke logging and screen monitoring software for up to 30 minutes at a time, following an internal revolt over the Model Capability Initiative that tracks wo…

OpenAI has published details of the Windows sandbox architecture powering its Codex coding agent, explaining that it built a custom isolation system after finding existing Windows security mechanisms insufficient for aut…

Anthropic warned Thursday that its AI models are advancing so rapidly they may soon be capable of improving themselves without human involvement, urging an industrywide pause on development. The AI lab, valued at nearly …

A computational cognitive neuroscientist who has studied alignment for three years is predicting that the first takeover-capable AI will be an advanced LLM augmented with human-like cognitive faculties such as continuous…

Staff-plus engineers are embedding cryptographic join keys (SHA, bundle hash, policy hash) into observability artifacts to quantify organizational leverage across AI/ML, Web Performance, and Cloud & Infrastructure domain…

Organizations adopting AI face growing security risks from Shadow AI and prompt injection vulnerabilities as AI shifts from tools to autonomous agents. The guidance provides practical steps to harden deployments and redu…

A startup, research lab, or hyperscaler announces a new tool to detect AI hallucinations every few weeks, marketing it as a "guardrail" for safe AI. However, any detector must already know the correct answer to catch a f…

Anthropic staff are embedded at the National Security Agency, helping the agency use the company's Mythos AI model for offensive cyberattacks, according to a Financial Times report. The model, which Anthropic deemed too …

Anthropic began red teaming its new Mythos model, codenamed Claude Oceanus-v1-p, on June 3, with early access granted to security testers. The preview candidate focuses on advanced reasoning, coding, and cybersecurity, b…