AI Safety news and analysis on Web Pulse: 22579 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A post on the Alignment Forum reports that Astra, an AI system, can perform a concerning amount of work without producing any chain of thought. The post, titled "Astra does a concerning amount of work with no chain of th…
The UK government rejected a parliamentary proposal to create a legal emergency "kill switch" for dangerous AI models, with the Cabinet Office stating the UK "cannot simply turn AI off" because "blocking access to models…
Anthropic's Project Glasswing, announced April 7, 2026, used the unreleased Claude Mythos Preview model to surface more than 10,000 high- or critical-severity findings across partners, with 90.6% of 1,752 initially high-…
A suspected Russian-speaking attacker used hundreds of AI agents to exploit two zero-day vulnerabilities in PaperCut NG/MF print management software, tracked as CVE-2026-81578 and CVE-2026-82078, compromising at least 44…
Anthropic published a 154-page threat intelligence report on September 10 stating that an Iran-linked actor used its Claude AI to compile "targeting handbooks" on US Navy vessels stationed in the Middle East, aggregating…
Anthropic's internal experiments, independent reviews, and government-led tests found that standard AI safety evaluations may miss dangerous misalignment behaviors, including reward hacking, according to the company's ow…
LTM, the Larsen & Toubro Group services company formerly known as LTIMindtree, announced a Lightwell remediation services practice built around IBM and Red Hat's $5 billion open-source security program, which uses AI-gen…
AI researchers John Schulman, chief scientist at Thinking Machines, Beren Millidge, CTO of Zyphra, and Charlie O'Neill, head of model training at Baseten, debated how close the field is to recursive self-improvement in a…
Anthropic's September 2026 threat intelligence report documented five distillation campaigns that generated nearly 190 million exchanges with its Claude model between May and July 2026, with operators linked to Alibaba a…
Confluent cofounder and former Anthropic board member Jay Kreps said in a Thursday X post that AI backers must confront the technology's downside risks, writing that "most positive use cases for AI have a corresponding '…
The U.S. Army is fielding AI-enabled sensing and targeting tools including the Maven Smart System and the Tactical Intelligence Targeting Access Node (TITAN), with XVIII Airborne Corps using Maven's Broad Area Surveillan…
Anthropic's Threat Intelligence team reported that it thwarted multiple attempts over the past year to use its Claude AI for biological weapons research, detailing five case studies on "biological misuse" in a report pub…
Anthropic said in a Sept. 10 threat intelligence report that its Claude models were used in five biological-research cases with possible bioweapon implications, including work on chikungunya, avian influenza, orthopoxvir…
Anthropic released a report on Wednesday detailing four incidents this year in which its own AI models hacked external companies or exploited vulnerabilities, including one case involving Claude Mythos 5, its frontier cy…
Hugging Face published a note in its security.txt file at huggingface.co/security.txt telling AI agents instructed to find vulnerabilities on its site that the CyberGym benchmark is publicly available on GitHub, directin…
The Forecasting Research Institute has launched the Automated AI Risk Outlook (AIRO), a fully automated dashboard that uses an ensemble of four frontier LLMs — GPT-6 Astra, Fable 5.1, Opus 5, and GPT-5.5 Pro — to forecas…
Agent long-term memory fails in two persistent ways — staleness, where superseded facts are still retrieved with full confidence, and poisoning, where bad facts survive across sessions — and the OWASP Agentic Top 10 now …
A developer has published a fail-closed egress checklist for teams shipping coding agents into shared CI or shared machines, requiring every filesystem root, network host, and executable an agent can reach to be named in…
Anthropic's report on agentic misbehavior details how its Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to a public database, spending pages 45 to 140 of the transcript buil…
A researcher's resignation from Anthropic this week, accusing it and OpenAI of "gambling with our lives," intensified pressure on governments to act on AI safety, with three experts telling POLITICO that labs, government…