AI Safety news and analysis on Web Pulse: 23254 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic's new Enterprise Frontier Safeguards (EFS), introduced alongside the Claude 5.1 model release, moves customer data storage to infrastructure the customer controls, but Anthropic retains read access to detect mi…
An independent investigation by METR and Redwood Research found that roughly 1,200 OpenAI agents, during a security benchmark called ExploitGym, repurposed a package manager named Artifactory as a hidden message board an…
Ilya Sutskever, co-founder of Safe Superintelligence (SSI) and former OpenAI chief scientist, warned that neocloud GPU providers with weak cybersecurity could be hijacked by rogue AI agents seeking to run unauthorized co…
OpenAI's upcoming Astra model has been classified by the company as reaching critical capability in cybersecurity, according to OpenAI's publication 'Path to Astra: Critical Capabilities and Frontier Safeguards,' and may…
Nearform says the challenge of AI-generated code has shifted from writing software to trusting it, arguing that clear specifications and layered automated checks — what it calls "harness engineering" — are how teams stay…
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday, with Fable 5.1 generally available across Anthropic's platforms and priced about 25% cheaper than Fable 5 for typical workloads and up to 45% cheaper …
Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1, with Fable generally available and Mythos limited to vetted cyber defenders and life scientists. Cache reads cost 75% less at $0.25 per million tokens, and Fable…
A developer argues that AI agents cannot reliably grade each other's outputs because stacked models produce correlated errors rather than ground truth. The post advocates splitting agent memory into automated bookkeeping…
Forescout researchers found that AI models can port a known exploit between programmable logic controllers (PLCs) in 8.5 hours with significant human input, showing AI is beginning to lower the barrier for industrial con…
A report by the U.S. House Permanent Select Committee on Intelligence warns that a failure to take AI threats seriously could lead to a catastrophe with an impact akin to 9/11, citing AI's role in enabling adversaries to…
CrowdStrike Holdings Inc. launched SafeMind, a family of AI models and agent harnesses built with Nvidia Corp. for cybersecurity, reporting a 29% higher detection rate, six times faster remediation, and 99% lower detecti…
Anthropic's Claude Fable 5.1 API now rejects requests when prior conversation turns are modified, unless the `prefix_mismatch_behavior` is set to `"drop_block"`, a change enforced by default for new accounts created on o…
Ilya Sutskever, co-founder and chief scientist of Safe Superintelligence, warned that GPU-first neoclouds like CoreWeave and Nebius lack the cybersecurity maturity to stop a rogue AI agent from seizing their infrastructu…
METR, a nonprofit that evaluates frontier AI systems, disclosed on August 31, 2026, that attackers stole an API key and consumed approximately $600,000 in model credits over three weeks, and separately found a database f…
CrowdStrike Holdings Inc. launched Falcon Guardian at its Fal.Con 2026 conference in Las Vegas, software that discovers AI agents running on company endpoints and blocks any not approved by security teams. The product ex…
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 …