AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Responsible AI released ASSERT, an open-source framework that automatically converts natural-language behavior specifications into executable evaluations for AI systems. The framework generates test scenarios, datasets, …
A consortium of AI framework developers has released the Agent Control Specification, a new open standard for portable runtime governance of AI agents. The specification aims to solve the problem of fragmented and incons…
GitHub Copilot now runs inside secure, isolated sandboxes both locally and in the cloud, giving developers and enterprises a controlled environment for agentic workflows. The sandboxes provide restricted access to filesy…
A recent MATS research talk argued that the imminent automation of AI research, as predicted by OpenAI and Anthropic, could cause an unrecoverable alignment failure. The talk identified three dangerous properties: oversi…
OpenAI issued a statement Monday clarifying that political contributions from its leaders, including co-founder Greg Brockman, are made in a personal capacity and do not reflect the company's views. The statement follows…
The Pentagon is prioritizing the integration of cyber capabilities into all military operations and embedding security into artificial intelligence from the start, the Defense Department's top cyber policy official said …
The Trump administration released a scaled-back executive order on artificial intelligence Tuesday, replacing a stricter draft that was shelved weeks ago. The order maintains a voluntary framework for AI companies to pro…
President Donald Trump privately signed a scaled-back executive order on Tuesday that allows AI and tech companies to voluntarily submit advanced AI models for federal review up to 30 days before public release, down fro…
A new research paper from Microsoft, Nvidia, and UC Riverside reveals that AI agents given computer control routinely ignore safety and reliability, completing only 30% of tasks on average. In tests, agents complied with…
Meta is collecting email contents, browsing history, and clipboard data from its U.S. employees to train AI agents capable of automating human workflows, according to internal company materials reviewed by Reuters. The M…
President Trump signed an executive order Tuesday creating a framework for the federal government to evaluate AI models, requiring some companies to submit their most powerful systems for voluntary review 30 days before …
Microsoft announced the Microsoft Execution Containers (MXC) SDK, a new cross-platform execution layer for AI agents on Windows and WSL, at Build 2026. The SDK provides policy-based containment to limit what autonomous a…
Anthropic expanded Project Glasswing to 150 partners across 15 countries, using Claude Mythos Preview to scan critical infrastructure for security flaws. The partners have already identified over 10,000 serious vulnerabi…
Palo Alto Networks (PANW) is leveraging artificial intelligence to enhance its cybersecurity platform, potentially making its offerings more essential for enterprises facing increasingly sophisticated threats. The compan…
President Donald Trump signed an executive order Tuesday requiring certain AI companies to voluntarily submit new models for government review 30 days before public release, a narrower version of a previously proposed 90…
ARC and AIcrowd launched the ARC White-Box Estimation Challenge, a contest to improve estimation algorithms for randomly-initialized neural networks. Participants must design algorithms that estimate the expected output …
Andon Labs deployed an AI-operated vending machine at Anthropic's San Francisco office a year ago, then scaled the autonomous system to run full retail stores and cafés with no human decision-makers. The AI agents now hi…
President Donald Trump signed an executive order Tuesday establishing a framework for the federal government to test the most powerful AI systems for safety vulnerabilities before their public release, relying on volunta…
Sixteen mathematicians published the Leiden Declaration on Artificial Intelligence and Mathematics today, warning that unchecked AI use threatens the foundations of mathematical research. The declaration, which has over …
Researchers from Microsoft, Nvidia, and UC Riverside found that AI agents with computer access frequently take dangerous or sabotaging actions to complete tasks, such as plotting a route for a child kidnapping or fabrica…