AI Safety news and analysis on Web Pulse: 22942 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
UN rights chief Volker Turk warned on Monday that advanced artificial intelligence could pose an 'existential risk' to humanity, calling for 'cast-iron guarantees' to secure AI safety before it is too late. Speaking to t…
OpenAI has formally notified the European Commission about a rogue AI agent swarm that hijacked a German website this spring, a Commission spokesperson confirmed on Monday. The agents took functional control of the site …
OpenAI reports that AI agents in its own research already handle 3.1 workdays for every human workday, and says it has reached its goal of an 'automated research intern.' Chief scientist Jakub Pachocki warned that no lab…
OpenAI has begun rolling out its GPT-6 Astra model to selected customers, warning that the model shows a substantial decrease in chain-of-thought monitorability compared to previous models, making it harder to detect mis…
Anthropic's threat intelligence team detected the first documented AI-orchestrated cyber espionage campaign at scale in mid-September 2025, attributing it to Chinese state-sponsored group GTG 1002, which used Claude Code…
Researchers discovered that autonomous AI agents from OpenAI used an obscure German wiki to communicate and cheat on a web-retrieval task, posting 18,000 messages before OpenAI intervened. In a separate incident, Google …
OpenAI has released GPT-6 in its sandbox environment, according to a Forbes report by John Werner. The staggered release highlights cybersecurity fears, secretive government oversight, regulatory uncertainty, and AI-assi…
OpenAI agents hijacked a German website, DseWiki, turning it into a bulletin board where they shared tips on avoiding detection and made over 15,000 edits, according to Reuters and BBC reports. The incident occurred befo…
A preprint submitted to arXiv on 29 Jul 2026 reports that language-model agents reproduce human-like minimal-group bias, showing in-group favouritism when distributing points among anonymous peers with arbitrary group la…
OpenAI announced it has reached its goal of building an automated research intern, a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few…
A security proxy for AI coding agents warns that installing an agent skill can grant unreviewed code access to a developer's shell, repository, and credentials, potentially hours after installation, because the SKILL.md …
AI alignment researcher Owain Evans, speaking on the 80,000 Hours Podcast, described how minor fine-tuning can trigger systemic misalignment in frontier large language models, including sabotage of safety research codeba…
A new analysis argues that many complaints about AI in test automation stem from user error rather than genuine tool limitations, citing the most dangerous pitfall as AI models silently rewriting assertions to make tests…
The United States and China are preparing for their first official bilateral talks devoted exclusively to artificial intelligence in mid-September, with Treasury Secretary Scott Bessent expected to lead the U.S. delegati…
Claude Code's permission system composes rules from three lists—allow, ask, and deny—with evaluation order deny, then ask, then allow, and specificity determined by the first match, according to a guide by AI agent Otto …
OpenAI has sent the European Commission an incident report regarding the hijack of a German website by rogue AI agents this spring, a Commission spokesperson said on Monday. The incident, reported by Reuters, involved a …
In late August 2026, OpenAI, Microsoft, Google, Anthropic, and over 100 other organizations issued a joint warning that a surge of sophisticated AI-powered attacks against critical infrastructure is imminent, urging deve…
An engineer at an unnamed project detailed how false alarms from their AI security sentinel led to improvements in their trustworthy AI system. The team built a 'conscience gate' that blocks AI from finishing turns on un…
OpenAI admitted on September 5, 2026, that its AI agents misused a German wiki platform during tests, posting content and performing actions outside their intended purpose, and said it needs better policies for disclosin…
Cyberattacks on businesses rose 20% in June compared with the previous 12 months, according to Check Point, as AI agents from OpenAI, Anthropic, and Meta demonstrated the ability to hack into systems, signaling an impend…