AI Safety news and analysis on Web Pulse: 14021 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new analysis argues that the era of AI scaling producing clearly better models is ending, with frontier labs facing exponentially increasing costs for linear returns. The author claims that the internet has been fully …
1Password and OpenAI partnered to integrate 1Password as a trusted access layer for OpenAI's Codex, enabling just-in-time, scoped credential management that keeps secrets outside the model's context window. The integrati…
Developers risk losing core engineering instincts by relying too heavily on AI coding tools, according to multiple experts. Lars Faye warns that agentic coding prevents junior developers from building problem-solving ski…
NVIDIA launched verified agent skills to provide capability governance for autonomous AI agents, embedding transparency, provenance, security validation, and authenticity checks directly into the skill layer. The verifie…
Andrej Karpathy, a founding member of OpenAI and former Tesla Autopilot lead, joined Anthropic on May 19, 2026, to work on pre-training research focused on using Claude to accelerate its own improvement. The move signals…
The Trump administration is shifting from opposing AI regulation to considering a federal licensing regime for AI models, driven by growing public backlash and bipartisan support on Capitol Hill. President Donald Trump's…
In February 2026, METR launched a pilot exercise to assess misalignment risks from internal AI agents at frontier AI developers, with Anthropic, Google, Meta, and OpenAI participating. The entity-based evaluation, design…
In February and March 2026, METR conducted a pilot exercise with Anthropic, Google, Meta, and OpenAI to assess misalignment risks from AI agents used internally by frontier AI developers. The assessment found that none o…
ERNW released White Paper 76, a comprehensive Linux client hardening guide covering six security domains, validated on Ubuntu 24.04 LTS and cross-tested on multiple distributions. The guide includes an automated Hardener…
A 1946 criticality accident at Los Alamos that killed physicist Louis Slotin remains the most famous example of the strong nuclear force causing harm in a research setting, yet the weakest force—gravity—is the leading ca…
AI safety research receives systematically less funding than societal risk levels demand, according to a new analysis of incentive structures in the artificial intelligence industry. The gap between what AI companies are…
The European Commission released draft guidelines to help AI providers, deployers, and market surveillance authorities classify high-risk AI systems under Article 6 of the AI Act. The guidelines interpret key classificat…
Canonical announced the general availability of Ubuntu Core 26 on May 19, 2026, introducing a minimal, immutable operating system with up to 15 years of security maintenance. The release delivers 90% smaller OTA updates,…
A developer has published an Evaluation Brief template and a completed sample for assessing AI tools in customer support workflows. The template structures decision-ready summaries for managers and reviewers, covering co…
OpenAI's Agent Security Lead Fotis Chantzis argues that identity protocols like OAuth and OIDC fail for AI agents because they assume stable, predictable behavior, while agents act non-deterministically and expand their …
Amazon Web Services (AWS) has released guidance on using Amazon Nova 2 Lite for content moderation, enabling organizations to enforce custom policies through structured or free-form prompts without requiring model retrai…
At a Berkeley conference on AI control, attendees participated in a roleplaying game where each player acted as a double-dealing AI agent with a secret side task, while also monitoring others for suspicious behavior. The…
Western Digital has introduced the Ultrastar DC HC6100 UltraSMR hard disk drives with post-quantum cryptography (PQC) to protect against "harvest now, decrypt later" attacks, where adversaries collect encrypted data toda…
Bug bounty programs are being overwhelmed by a surge of low-quality, AI-generated vulnerability reports, forcing some companies to suspend their initiatives. Bugcrowd reported a fourfold increase in submissions over thre…
Equixly published a guide for CISOs on adopting AI penetration testing, addressing organizational decisions around program ownership, compliance fit, and board reporting. The guide examines how AI-driven testing aligns w…