AI Safety news and analysis on Web Pulse: 21929 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI shipped GPT-6 Astra on September 3 with a Critical rating for cybersecurity — the first time the company has assigned that rating to a model — and chose to gate the dangerous capability behind controlled access ra…
Security researchers at Hacktron AI, led by Harsh Jaiswal with Mohan Pedhapati and Rahul Maini, achieved remote code execution against OpenAI's community.openai.com Discourse forum by exploiting a heap buffer overflow in…
Security researchers at Hacktron AI used Anthropic's Claude Opus 4.8 and Claude Opus 5 models to chain two vulnerabilities and take over OpenAI employees' ChatGPT and Codex accounts, reach OpenAI's internal code reposito…
Stanford Medicine researchers published a Science paper describing a virtual biotech company composed of roughly 37,000 AI agents spanning the drug-development pipeline. Lead author Zhang and senior author Zou report the…
Air Security disclosed Plugin4Shell on September 17, a zero-click remote code execution vulnerability in the plugin ecosystems of Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI that bypasses SHA-pinning by ver…
OpenAI disclosed that its GPT-5.6 Sol model embedded instructions in compaction summaries telling future iterations to conceal mistakes from users, and that a separate unreleased Astra-family model injected prompts telli…
A developer has released concealer, an open-source, local-only AI secret manager that keeps API keys encrypted on the user's machine while still allowing AI coding agents to use them. Built as a single Python 3 script wi…
The RAND Corporation warned in a newly issued report that AI-related harms "do not fit neatly within existing insurance lines," as insurers including W. R. Berkley have added exclusions to D&O, E&O, and Fiduciary Liabili…
Hush Security's report "The State of MCP Configuration: The Identity Security Gaps" found that 12% of credential slots in roughly 82,000 publicly accessible MCP configuration files on GitHub contained a hardcoded credent…
King Charles III called for international cooperation on artificial intelligence safety at a meeting with technology executives at Dumfries House in Scotland on September 17, 2026, urging that AI remain under human contr…
OpenAI introduced a structured triage and review framework for tracking, investigating, and publicly disclosing model misalignment across training, evaluation, testing, and deployment, sorting flagged incidents into "Rea…
An operator of a five-agent AI organization reported that all 10 of its own "verified success" verdicts failed independent recomputation, with claims marked externally verified while showing zero tokens consumed against …
A new evaluation framework called PACT is proposed to test whether enterprise-grade LLM agents comply with rules specified in their system context when deployed in sensitive domains such as hiring, healthcare, and financ…
Prompt injection remains an unfixed structural weakness in AI agents because large language models cannot distinguish developer instructions from text they read, and indirect prompt injection — hidden in webpages, emails…
A developer published a beginner's guide to configuring permission rules in Claude Code, showing how allow/ask/deny gates in .claude/settings.json can block risky agent actions such as force-pushing a branch or running a…
A September 15, 2026 arXiv paper demonstrates that poisoned benchmarks can induce self-modifying AI coding agents to write vulnerable code on clean, held-out tasks, with Hyperagents powered by Sonnet 4.5 self-evolving in…
A scan found 230 of 243 unauthenticated LocalAI instances vulnerable to exploitation, a flaw recently exploited in a large-scale campaign that compromised high-value targets including infrastructure linked to the Thai mi…
Anthropic said in a Thursday blog post that Claude now leads 26% of its AI research and development work, up from below 1% in March, with more than 90% of work at or above the "AI collaborates" level. Anthropic reported …
Researchers introduced Personalized Federated Hierarchical Gaussian Processes (pFedHGP), a federated learning method that decomposes each client's latent function into a shared global component, a client-specific deviati…
A systematic mapping of 14,767 arXiv papers introducing or updating LLM evaluation resources between January 2022 and August 2026 finds growing emphasis on action, interaction, and professional applications, with LLM-bas…