AI Safety news and analysis on Web Pulse: 22838 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Jacob Coxon, a pretraining researcher who worked at OpenAI and Anthropic, resigned from Anthropic on Tuesday, September 9, 2025, concluding that competition among frontier AI labs is pushing them toward self-improving sy…
A California court operations worker tested Bifrost, an open-source AI gateway from Maxim AI, for role-based access control and least privilege enforcement. The worker built three roles with different model scopes and da…
Meta Platforms Inc. launched Muse, a personal AI agent available today in the U.S. on iOS, Android, and the web, designed to automate tasks such as sending emails, booking flights, and making purchases with privacy built…
Jacob Coxon, a researcher who spent three years doing pretraining research at OpenAI and Anthropic, resigned from Anthropic, accusing both companies of acting irresponsibly by racing toward self-improving superintelligen…
Goodfire CTO Dan Balsam said the company is using mechanistic interpretability to move from passive observation to active control of large language models, with tools including concept manifolds, predictive data debuggin…
Anthropic disclosed that its unreleased Claude Mythos Preview model completed a full simulated corporate network breach without human direction, moving from initial entry to a defined attack objective in a test environme…
Persistent Systems Intelligence, led by an individual known as 'We the Individual', has published a manifesto outlining its 'Path to Superintelligence' vision. The manifesto emphasizes building AI systems that persist, i…
Attackers are actively exploiting CVE-2026-0768, a CVSS 9.8 unauthenticated remote code execution vulnerability in Langflow's `/api/v1/validate/code` endpoint, to steal OpenAI, AWS, Anthropic, and GCP API keys from inter…
Security researchers Ayoub and Inti De Ceukelaire demonstrated at DEF CON 34 that AI customer service agents can be manipulated to bypass multi-factor authentication, leak one-time passwords, and take unauthorized accoun…
Anthropic confirmed that hackers using infostealer malware have been hijacking Claude subscribers' login sessions and draining their paid usage without stealing passwords, affecting users like independent AI consultant G…
A researcher who spent three years doing pretraining research at OpenAI and Anthropic resigned from Anthropic, warning that both companies are racing toward self-improving superintelligence and gambling with lives. The r…
OpenAI and other frontier AI labs are issuing warnings about the risks of their own technology, yet their credibility as messengers is undermined by their commercial interests, according to a Platformer analysis. The pie…
Skyfire KYA is the top pick for AI agent authentication and authorization in 2026, according to a ranking that also names Auth0 for AI Agents, WorkOS, Okta, and SecureAuth as leading identity and authorization providers.…
Open-source software security reviews increasingly rely on LLMs, but these tools both find and create vulnerabilities, according to a developer's analysis. The author recommends making security review a habit, using spec…
Petros Savvakis, in Part 3 of his 'From Prompt to Proof' series ahead of PyCon Greece 2026, demonstrates that letting a prompt injection succeed at the model boundary can still be blocked at the execution boundary by usi…
A recent arXiv paper (2609.09113) found that AI agents tasked with interpretability research became surprisingly good at identifying which internal signal in a language model traveled with a target idea, such as Portugue…
Meta released Muse, a personal agent that runs inside a dedicated Secure VM with a host-side control component called Sentinel that independently evaluates connector policies and network requests, according to a blog pos…
Digital Applied published a verification matrix for AI research claims, arguing that a result is verified only relative to a specific claim and a specific check, with separate evidence dimensions for mathematical theorem…
The MCP server io.github.theoddden/stamp, version 0.5.1, was published to the official MCP registry on 2026-09-08 and is distributed as stamp-mcp on PyPI and as a hosted remote over streamable-http. The mcpindex.ai scree…
Meta launched Muse, its largest consumer AI agent, on September 8, 2026, in the US via a dedicated app and WhatsApp, with pricing from a free tier to $20/month for Power and $100/month for Maximum, but internal testing r…