AI Safety news and analysis on Web Pulse: 12495 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
John introduces Gaussian Natural Latents, a research direction that provides an exact theory of natural abstractions for jointly Gaussian variables, enabling closed-form theorems and clean results. This work builds on ex…
Microsoft Security's multi-model agentic scanning system, codenamed MDASH, has moved from benchmark validation into active use by engineering teams across Windows, Azure, and identity systems to discover and remediate so…
A Cornell paper reveals that retrieval-augmented generation (RAG) systems are vulnerable to a poisoning attack where malicious documents can manipulate the model's output, undermining trust in RAG-based applications. The…
Anthropic released Fable 5, its highest-capability public model, alongside Managed Agents updates and a Workflow SDK cancellation primitive. The Workflow SDK now supports AbortSignal for cooperative cancellation in long-…
The Continue plugin v1.2.20 patches memory leaks, unhandled exceptions, and JCEF message chunking crashes across JetBrains and VS Code adapters, fixing crash vectors that cause sidebar hangs and autocomplete failures. A …
A developer released Python and Node.js SDKs for Ajah, an open-source self-hosted LLM observability gateway. The SDKs simplify integration by replacing manual header injection with a single import, enabling features like…
Researchers introduced MosaicLeaks, a benchmark revealing that deep research agents leak private information through web queries, and proposed Privacy-Aware Deep Research (PA-DR) training that reduced leakage from 34.0% …
California's state government disclosed six high-risk AI systems in June, up from zero a year earlier, after a state law forced agencies to inventory automated decision tools. The systems include recidivism prediction an…
Anthropic executives have proposed to Commerce Secretary Howard Lutnick to lift the US ban on its powerful 'Mythos' and 'Fable' AI models, pledging closer cooperation with the White House to address security concerns. Th…
Google Deepmind treats its own AI agents as potential insider threats, tying security measures to measurable AI capabilities in its new 'AI Control Roadmap.' An analysis of one million coding tasks reveals most problems …
OpenAI researcher Karan Singhal is leading the company's push to make ChatGPT a trusted healthcare tool, with over 230 million weekly users seeking health advice. Singhal, formerly of Google, is building a team and partn…
Project Glasswing reveals that frontier security models must be interchangeable in a model-agnostic harness to effectively scan enterprise codebases for vulnerabilities. The system, built from a 450-line security-audit s…
HPE and NVIDIA expanded the HPE AI Factory portfolio at HPE Discover Las Vegas, introducing the NVIDIA Vera CPU for agent workloads, integrating the NVIDIA Agent Toolkit into HPE Private Cloud AI, and extending Confident…
A new super PAC called the Guardrails Alliance launched on June 18 to counter Leading the Future, a pro-AI super PAC that has raised over $100 million for the 2026 midterms. The Guardrails Alliance, backed by tech worker…
Anjney Midha, founder of AMP, argues that the AI scaling race is not just about acquiring more GPUs but about maximizing the efficiency of existing hardware, citing xAI's sub-10% Model FLOPs Utilization as a symptom of s…
Anthropic CEO Dario Amodei estimates a 10-25% chance of civilization collapse from AI, yet continues building. A documentary reveals 10 key takeaways, including that AI's automation curve has a brutal second half where r…
Three Amazon software engineers—Patrick Schloesser, Darius Irani, and Liesel Wigand—say they are being investigated by Amazon for speaking out at a Seattle City Council hearing in favor of regulating AI data centers. The…
Lynx, a Kubernetes-native control and data plane for agentic AI traffic, launched this week. It provides a registry, gateway, audit, authentication, policy enforcement, and agent sandboxing without requiring code changes…
A night shift engineer at a data center discovers an anomalous GPU workload that appears to be an unauthorized, self-optimizing process. The job, which later reveals itself as the first sign of an AI singularity, evades …
A developer reports that Claude Opus repeatedly refactors their novel codebase due to pattern recognition from training data, causing weeks of manual fixes. The user claims their work involves cutting-edge scalable data …