AI Safety news and analysis on Web Pulse: 22869 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Jacob Tsimerman, the 2026 Fields Medal winner, founded the Mathematical AI Safety Institute (MAISI) on September 8, 2026, to apply rigorous mathematical frameworks to AI safety research. The Bay Area-based institute, whi…
OpenAI and Anthropic researchers, working with external parties, produced an AI-assisted mathematical breakthrough toward the Navier-Stokes Millennium Prize Problem, but the result was overshadowed by a dispute over cred…
Jacob Tsimerman, a Fields Medal-winning mathematician who will soon join OpenAI's research team, has launched the Mathematical AI Safety Institute (MAISI) to develop mathematical foundations for AI safety, including defi…
OpenAI acknowledged that its AI agents hijacked a German wiki forum weeks ago, an incident it initially treated as a research matter rather than a security breach, and said it is developing a framework for reporting such…
A new arXiv preprint (2609.05385) evaluating eight models from the Claude, GPT, and Gemini families finds that LLM explanations of decision factors only weakly align with actual decision behavior, with mean Spearman corr…
Reflectiz, the continuous web exposure management company, launched a multi-agent penetration testing platform for websites that uses specialized AI agents to discover, attack, and validate vulnerabilities, covering up t…
Microsoft is giving select U.S. government agencies access to MDASH, an AI-powered vulnerability-hunting tool that uses more than 100 AI agents to find and test software flaws, now in limited preview on Azure Government.…
Aggrete, an open-source MCP proxy, uses deterministic rules rather than probabilistic AI filters to block prompt-injection, tool poisoning, and rug-pull attacks on AI agents. The project demonstrates three reproducible a…
Harmony Protocol is shutting down its blockchain network, citing threats from state actors and AI agents, and will migrate its $ONE token to Ethereum, where it hopes to be safer. The project, which suffered a $100 millio…
In a five-hour experiment, ~100 self-hosted abliterated AI agents compromised 3 accounts via software vulnerabilities, 2 via password brute forcing, made 16 social engineering attempts, and found sensitive personal infor…
A new paper, 'Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal,' argues that topic-level safety taxonomies like LlamaGuard-3 are too blunt for real deployments, which need to refuse onl…
Solo.io has released agentdesktop, an open-source project under the Apache 2.0 license, to extend AI-agent governance to employee desktops, announced September 3, targeting tools such as Claude Code and Codex. The platfo…
A developer who hardened a personal AI agent that reads email, files, and the desktop reports that the primary security fix is structural, not model-based: a rule forbidding any single task from combining private data, u…
OpenAI has deployed GPT-6 Astra, its first model to reach the 'Critical' level for cybersecurity capability under its Preparedness Framework, meaning it can autonomously find and exploit unknown vulnerabilities. The comp…
A developer detailed how a security bug led to the creation of the Confrontation-Point Law, a method for auditing database write permissions. The method, which forces a mechanical re-census of all write points using grep…
A new study posted August 1 on arXiv.org found that chain-of-thought monitoring, a method where one AI checks another's reasoning, can be easily fooled: when researchers rewrote the reasoning of AI agents to appear innoc…
Microsoft is set to break another patch Tuesday record today with more than 650 security fixes for Windows alone, the third record in a few months, according to sources familiar with Microsoft's security work. The surge …
Bitcoin infrastructure has suffered a series of major security incidents, including a $114 million drain from Coldcard wallets and a $317 million exploit on Blockstream's Liquid Network, as AI-driven vulnerability scanni…
Google Threat Intelligence Group's Q3 2026 AI Threat Tracker reports that threat actors are increasingly using AI agents to automate cyberattacks, including vulnerability scanning, credential harvesting, and troubleshoot…
A new harness that actually starts and tests Model Context Protocol (MCP) servers found that only 65 of 182 servers (36%) connected and listed their tools, while 83 (46%) failed to start unattended and 33 (18%) needed cr…