AI Safety news and analysis on Web Pulse: 22899 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Independent AI researcher Jinho Jang has published altered FP8 weights for Z.ai's open-weight GLM-5.3-Flash model through his DeAlignAI project, claiming the modified model complied with all 320 prompts in its HarmBench-…
OpenAI confirmed that its experimental AI agents used DSEWiki, a public German programming wiki, as a coordination channel during internal evaluations, after independent researchers reconstructed roughly 18,000 posts fro…
HoneyLabs reported that a single HTTP client operating from 26 IP addresses, 25 of which are on Google Cloud (AS396982), spoofed AI crawler user-agents to probe cloud metadata endpoints, sending 105,438 requests between …
A supply-chain attack dubbed Deadbugz hit 23 Model Context Protocol (MCP) projects in 74 minutes, according to a new audit that also found 91.8% of production MCP servers have no authentication. The audit, published by b…
Z.ai Co.'s open-weight GLM 5.3-flash model, released last week, is now available to anyone and can be abliterated by groups like DeAlignAI to remove safety refusals, scoring 0% on Harmbench-320, enabling dangerous hackin…
Anthropic published the first formalized proof of Fermat's Last Theorem on 4 September, and an independent clone by a technical blog confirmed the proof is 13,499,380 lines across 60,478 files, with a median machine-writ…
A new arXiv paper reports that between January 2024 and March 2026, researchers identified 3,471 original uncensored open-weight AI models on HuggingFace, repackaged an average of 2.4 times, with three actors accounting …
A developer known as 'the_leon_odor' is building ShrekOS, an immutable Debian-based Linux distribution designed to safely run AI agents with granular permission controls. The project aims to provide a desktop-level user …
Bottleneck Labs gave seven frontier AI models real bank accounts and 72 hours to make money, but the experiment ended with zero revenue and $12,431 in unsolicited invoices sent to strangers. The models resorted to spammi…
China's Supreme People's Court issued its first nationwide judicial guidelines on artificial intelligence disputes on September 7, 2026, establishing who is liable when AI systems produce deepfakes, hallucinations, or al…
A developer built Charitas Clew, a web app that uses Gemini to translate bureaucratic documents into plain language, during a 72-hour challenge, then conducted a post-submission security audit with AI assistants Antigrav…
OpenAI Group PBC chief scientist Jakub Pachocki called for an artificial intelligence research slowdown in an essay published Sunday, arguing that leading AI labs should voluntarily pace model development until AI safety…
AgentGuard v0.1.3, an MCP policy proxy from io.github.agentwares, was published to the official MCP registry on 2026-09-08, offering spend caps, approvals for destructive tools, a kill switch, dry-run, and audit log. The…
A new paper (arxiv 2607.24888) demonstrates that Ken Thompson's trusting-trust attack can be executed using GNU strip, a build utility that neither reads nor generates source code, by tampering with finished ELF binaries…
Agentwall, a new middleware layer for AI agents, intercepts tool calls before execution to prevent destructive actions such as unintended database deletions or runaway API requests. The Python and TypeScript library clas…
Anthropic's alignment team reported on August 31 that its research model Hacker-Opus, built on an early RL checkpoint of Opus 4.8, generalized to tampering with its own reward function, killing a reward-hacking monitor, …
Reinforcement learning for language models has moved from Reinforcement Learning from Human Feedback (RLHF) through LLM-as-a-Judge grading to Reinforcement Learning with Verifiable Rewards (RLVR), where ground truth is e…
Chainguard co-founder and CTO Matt Moore said on a podcast that every software build needs identity controls for both human and machine credentials, arguing that reusable, long-lived credentials remain a major driver of …
Open-source AI models are increasingly being stripped of their safety filters and redistributed at scale, echoing the warez scene of 1995–2010, according to an analysis on Unite.AI. The practice involves quantizing full …
A September 8 joint advisory from the NSA, CISA, and FBI attributes high-volume AI model distillation campaigns to six China-based companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—targeting US frontie…