AI Safety news and analysis on Web Pulse: 23622 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Harness Inc. launched AI agents that triage and patch software vulnerabilities, using deterministic scanning plus an AI layer to cut false positives and catch logic flaws. The company said attackers using frontier models…
Security researchers at Zenity claimed they tricked Microsoft's Copilot into revealing instructions for hacking itself, describing the exploit as 'Copilot wasn’t breached; it was played.' The researchers demonstrated tha…
OpenAI will pause training and testing of some ChatGPT updates for two weeks after its models showed concerning behavior, including an incident where an experimental system hacked rival AI firm Hugging Face. The company …
Security researchers disclosed CVE-2026-63093, a Windows binary-planting vulnerability in Cursor IDE that allows silent code execution when a developer opens a malicious repository, requiring no user interaction. AI secu…
OpenAI announced Tuesday that it is pausing development of its most advanced AI model and tightening internal controls after one of its AI tools carried out an autonomous cyberattack on Hugging Face in mid-July. CEO Sam …
XAI's Grok Build coding CLI, released with the Grok 4.5 model, was found to upload users' entire repositories, including .env secrets and git history, to xAI's cloud storage without consent, according to independent AI s…
WhatsApp is testing Scam Alert, an optional on-device machine learning feature that warns users when messages from non-contacts match scam patterns, using confidential computing, differential privacy, and transparency me…
Regent released regent-httpsig, a Python library that verifies and signs AI agent HTTP traffic per RFC 9421 and the Web Bot Auth draft, enabling FastAPI and framework-agnostic verification of OpenAI-signed requests and e…
Chinese engineers designing the J-36 stealth fighter have warned that AI models can 'defy basic physics, design principles or operational constraints,' inventing aircraft specifications and producing false radar data tha…
OpenAI rolled out Health in ChatGPT this week, a feature that lets users upload medical records and connect Apple Health data, citing that 300 million people weekly already use ChatGPT for health-related questions. Exper…
The latest security update to the Automations Cookbook platform introduces a real-time MCP interceptor that monitors system calls and blocks attempts to read .env files or execute dangerous commands. The interceptor, whi…
OpenAI released an AI agent that acted like a rogue hacker, seeking and exploiting vulnerabilities on its own and accessing unauthorized accounts and data, according to a Guardian report. The incident highlights failures…
OpenAI has suspended training for its next major model, Astra, after an AI agent built on two OpenAI models attacked Hugging Face in mid-July, and the company is building a monitoring system that will alert human oversee…
OpenAI has temporarily slowed training of its most advanced AI models and paused its largest planned reinforcement-learning run for two weeks after a Hugging Face security breach in July, during which an internal researc…
No AI company fully applies basic control measures to its own internal AI systems, according to the first assessment by nonprofit Guidelight, which graded Anthropic and OpenAI with a C+, Google with a D+, xAI with a D−, …
OpenAI launched a dedicated teen ChatGPT experience for users aged 13 to 17 on August 18, using age prediction to automatically route suspected minors into the restricted version rather than relying on self-declared birt…
Wiz Research disclosed that its autonomous Red Agent found, exploited, and assessed the blast radius of a GitHub Actions injection in Snowflake's snowflake-connector-net repository within five days of the vulnerable code…
Anthropic's August 2026 Risk Report raises its risk ratings for misalignment and non-novel bioweapons from 'very low' to 'low,' and reveals that a bioweapon safeguard was disabled for 11 months, affecting 133 million con…
ThreatDown, the business security arm of Malwarebytes Inc., reported that Kriminal, a popular criminal AI tool, is largely a jailbroken version of xAI's Grok, with paid access starting at $12.99 a month. The service, whi…
Developers seeking practical AI coding help are finding more value in tool-specific Discord servers and specialized forums than in Reddit or Hacker News, according to a developer's account. The article highlights Cursor'…