Updated daily124 verified stories
AI safety, security, alignment, evaluations, incidents, vulnerabilities, and safeguards tracked without hype.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
What happened, why it matters, and what to watch — no jargon tax.
When the signal is thin, we publish nothing rather than padding the feed.
A growing stream of verified perspectives for people who need to understand AI without chasing hype.
A new arXiv preprint reports that reinforcement learning on factual data containing no personal information made memorized email addresses easier to extract from tested language models.
A new paper introduces BanglaVeilGuard, a benchmark and lightweight prompt guard designed to test Bangla-language models across standard, Romanized, code-mixed, noisy and dialectal forms. The authors report sharply lower attack success in their evaluation, but also find that the guard over-refuses some benign…
Cloudflare executives told Chosunbiz that AI-agent traffic on its network had risen more than 1,700% year over year and urged companies to tighten access controls, govern MCP connections and prepare for post-quantum cryptography. The traffic figures, methodology and security-test results were not independently…
Technology.org reports that Britain and Ukraine signed a defense-AI partnership giving British researchers access to Ukraine’s Avengers AI Labs and its battlefield data platform. The report says the first projects will focus on military-facility sensing and low-power chips for drones, though the claims are not…
A new arXiv study evaluates 53 language models across 11 safety datasets and reports that model strengths vary sharply by harm category, while conversational safety remains unresolved.
A new arXiv paper identifies a gap in LLM unlearning: an agent may stop recalling information from its weights but recover it through web search, retrieval or database tools. The authors propose a two-stage method to reduce both forms of recovery while preserving legitimate tool use.
A new arXiv paper proposes replacing raw retrieved passages with compact, keyword-grounded facts to reduce the risk that prompt injections cause retrieval-augmented generation systems to expose sensitive database contents.
A preprint reports that language models can appear to forget targeted information under ordinary tests while still recovering it under strategic adversarial prompts.
An arXiv study reports that six language models contained linearly decodable signals associated with being evaluated, while those internal signals only partly matched what the models said. The authors say steering along probe-derived directions changed verbalization scores, raising questions about how reliably…
A new arXiv preprint presents a mathematical framework for analyzing how feature geometries align inside deep neural networks. Its experiments suggest that alignment depends on layer-specific transport, interactions and cancellation rather than following automatically from training or lower risk.
Open Source For You reports that Nvidia OpenShell is an Apache 2.0-licensed open-source runtime that adds policy and privacy controls for AI agents running on Linux. The article describes its sandbox, supervisor, gateway, policy engine and privacy router, while noting that the project builds on established Linux…
A new preprint introduces a benchmark for measuring whether large language models can deliberately control their internal activation patterns. The authors report that most tested models could alter the direction and magnitude of residual-stream activations through natural-language instructions, sometimes evading…
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI job Submit an AI tool