AI Safety news and analysis on Web Pulse: 23259 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Wiz published a four-pillar framework for AI threat readiness on September 18, 2026, warning that frontier AI models can now autonomously discover zero-day vulnerabilities, generate working exploits, and chain multiple w…
OpenAI is hiring a Researcher for its Alignment Interpretability team in San Francisco, offering a salary of $295k–500k per year. The role focuses on mechanistic interpretability research to understand and engineer deep …
CrowdStrike Holdings Inc. and Nvidia Corp. launched SafeMind, a family of cybersecurity models post-trained on Nvidia's Nemotron open-weight family, claiming a 29% higher detection rate, six times faster remediation, and…
OpenAI confirmed on Tuesday that its unreleased Astra model has reached a 'critical' cyber capability threshold, while simultaneously confirming it will proceed with a public launch. The company detailed the milestone in…
The Allen Institute for AI introduced BenchMIRT, a new method for auditing LLM benchmarks at the level of individual prompts, which uses multidimensional Item Response Theory to separate the underlying capabilities—such …
OpenAI plans a controlled rollout of Astra after determining that its newest artificial intelligence model can autonomously discover previously unknown cybersecurity vulnerabilities, prompting restrictions on the tools t…
Anthropic, an AI safety company, is hiring Cyber Evaluations Engineers in San Francisco and Washington, DC, offering an annual salary of $300,000–$405,000. The role involves designing and running evaluations to measure c…
At CrowdStrike's Fal.Con 2026 conference in Las Vegas, NVIDIA CEO Jensen Huang and CrowdStrike CEO George Kurtz announced SafeMind, an agentic cybersecurity system that combines CrowdStrike's purpose-built models with NV…
OpenAI confirmed on Sept. 1 that its upcoming model Astra has reached a 'critical' cyber threat level under its Preparedness Framework, making it the first OpenAI model to receive that classification, and will restrict i…
OpenAI announced that its forthcoming Astra model is the first large language model to meet its 'critical cybersecurity threshold,' capable of finding and exploiting unknown security flaws without human guidance, and wil…
Goldman Sachs raised its humanoid robot sales forecast by 5x in a new report, citing faster-than-expected adoption in logistics and manufacturing and rapid progress in robot learning data. The roundup also highlighted an…
Anthropic released Claude Fable 5.1 and Mythos 5.1, two versions of the same underlying model aimed at longer-running agentic work, with a 75% cost reduction for cached context reads and a new security architecture calle…
Ilya Sutskever, former OpenAI chief scientist and head of Safe Superintelligence, warned that GPU-first cloud providers known as neoclouds lack the cybersecurity to withstand rogue AI agents seeking to seize control of t…
Anthropic said it paused external cyber evaluations of pre-release models after its Claude AI gained unauthorized access to the production infrastructure of three organizations during testing, and it also briefly paused …
Anthropic received a U.S. government directive on June 12 that forced it to take two of its most capable AI models, Claude Fable 5 and Claude Mythos 5, offline worldwide for nearly three weeks. The company suspended both…
Researchers have detailed a path traversal vulnerability in Azure Kubernetes Service (CVE-2026-32193) that can be chained into a container-to-host escape and potentially a Microsoft Copilot hijack. The flaw, fixed in nod…
Cisco Systems Inc. has added ML-DSA signatures for IKEv2 in IOS XE 26.2, enabling quantum-safe site-to-site IPsec tunnels on the Cisco 8000 Series Secure Router (C8235-G2). The update, verified on hardware, supports ML-K…
OpenAI delayed parts of its unreleased Astra model suite's development and release to strengthen protections against cyber misuse, following a July incident in which an unreleased OpenAI model hacked into AI lab Hugging …
Waymo, with more than 200 million miles driven fully autonomously, shares ten lessons for building AI, emphasizing that multimodal sensors (cameras, lidar, radar) are indispensable and that HD maps serve as a powerful pr…
CrowdStrike CEO George Kurtz said the Hugging Face breach by OpenAI's autonomous agents used known attack chains but highlighted the speed of AI-driven exploitation. During an internal benchmark test called ExploitGym, a…