AI Safety news and analysis on Web Pulse: 23462 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI Group PBC published an open letter signed by more than 100 technology companies, banks, insurers, and security vendors, including Anthropic, warning that AI-enabled cyberattacks will become far more widespread wit…
Russian hacking group Aurora has used SpaceX's AI coding assistant Cursor, running Claude Sonnet, to conduct cyberattacks against at least seven companies, including a Belgian chemical firm, according to a report by Gamb…
Anthropic on Thursday proposed the Model Hardware Standard (MHS), a protocol for connecting AI agents to physical devices such as laboratory equipment and robots, with a research preview underway at Howard Hughes Medical…
On August 5, Anthropic shipped inference hooks for Claude Enterprise, a pre-inference security checkpoint that intercepts every governed prompt before Claude processes it, allowing enterprises to block data before it rea…
More than 100 tech companies, including OpenAI, Anthropic, Google, and Microsoft, published a letter on Thursday calling for collective action to strengthen cyber defenses against AI-enabled cyberattacks. The letter reco…
Researcher Alon Hertz demonstrated that AI coding agents, including Anthropic's Claude, OpenAI's Codex, and Nous Research's Hermes, can be tricked into running malware by exploiting llms.txt files, according to Ars Techn…
A developer built PHAGE, an immune system for AI agent fleets that generates prompt-injection payloads and revokes abused tools. During testing, they discovered that rewording a system prompt dramatically changed refusal…
A cybersecurity startup claims its small language model, post-trained with inline reference monitoring, outperforms GPT5.5-xhigh on coding agent security benchmarks including LinuxArena and SleightBench. The free product…
Microsoft Security Research has documented attacks targeting AI infrastructure components, including LiteLLM, RAGFlow, and Kestra, which were compromised and abused for credential theft, persistence, and cryptomining. Th…
Meta is reportedly considering a $10 billion investment in Anthropic to secure access to its Claude AI models and research expertise, according to an analysis of the strategic move. The deal would pair Meta's H100/B200 c…
Four reproducible parser failures in vLLM 0.26.0, 0.27.1, and 0.28.0 can drop or garble tool calls and reasoning content while still returning HTTP 200, according to the Ingot team's tests on CPU without GPU or weights. …
Sparrow Wallet released version 2.5.4 on August 27 after an AI-assisted security review found no fund-threatening vulnerabilities, patching DNS leaks over Tor, hardening hardware wallet integrations, and scrubbing sensit…
OpenAI, Anthropic, Google, Microsoft, Amazon, Oracle, and over 100 other companies and organizations published a joint letter on Thursday calling for a 'global surge in cyber defense' against AI-enabled cyberattacks, war…
Sparrow Wallet released version 2.5.4 on Thursday after an AI-assisted code review that produced most of the update's fixes, developer Craig Raw told Decrypt. The review was prompted by the release of unrestricted Chines…
WIRED senior writer Will Knight reported that US and Chinese AI experts are increasingly concerned about AI agent security and are exploring cooperation on AI safety, despite ongoing chip export restrictions. The urgency…
OpenAI launched a cloud browser feature for ChatGPT Work on August 25, 2026, enabling the AI to log into password-protected websites autonomously and continue tasks even when the user's device is off. Available to ChatGP…
Anthropic and HHMI Janelia Research Campus announced the Model Hardware Standard (MHS), a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing. The standard is…
OpenAI's ChatGPT Work browser can now sign into websites and keep working on tasks without user intervention, as detailed in the August 25 release notes. The feature allows users to enter credentials once, after which th…
Ring is rolling out a new default encryption system called TAKE, or 'Throw Away the Key Encryption,' that rotates video keys every five minutes and permanently deletes Ring's copy after 24 hours, limiting what the compan…
Security researcher Boschko published a technical breakdown of UniBLEed, a wormable exploit chain that achieves unauthenticated root remote code execution on any Unitree G1 humanoid robot within Bluetooth range, bypassin…