AI Safety news and analysis on Web Pulse: 23594 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A zero-click remote code execution flaw in Cursor's AI coding agent, tracked as CVE-2026-26268, allows attackers to execute arbitrary code via a malicious Git repository. Moonshot AI's open-weight Kimi K3 model ranks thi…
A developer working on agentic AI for production support is advocating for treating LLMs as a data egress boundary, proposing a sanitization layer before data reaches RAG or embedding models. The approach emphasizes reda…
A new paper on arXiv (2608.19215v1) proposes shifting from intractable questions of AI consciousness to tractable questions of AI valence to guide ethical treatment of potentially sentient AI. The author argues that asse…
A new arXiv paper (2608.19216v1) introduces 'bounded sovereignty' to describe the partial technical and contractual access deployers have over frontier AI models they do not own, and proposes a 'sovereignty discount cost…
A new framework called Symposium, described in an arXiv paper (arXiv:2608.19511v1), provides long-term, immutable records of AI agent-driven research for small scientific communities, enabling auditable trails and trust …
Researchers introduced a black-box method that classifies unsafe LLM outputs by fitting Koopman-based predictive models to prompt-response embedding dynamics, using a differential residual score to compare safe and unsaf…
A new arXiv paper (2608.19369v1) proves that statistical watermarks for language models are eroded by meaning-preserving transformations in a way that depends on the location of edits, not just their endpoint. The author…
Airbus and its partners have developed a supervised machine learning model that estimates helicopter weight during takeoff using data from Airbus's global in-service fleet, and demonstrated its implementation on legacy a…
A new study from arXiv (2608.19351v1) systematically evaluates template-based proof sharing for neural network robustness verification, finding that template subsumption rates vary widely across architectures, properties…
NVIDIA, Protect AI, Lakera, and other vendors offer AI agent security frameworks designed to defend against prompt injections and tool hijacking in 2026. These tools provide multi-layered defenses including programmable …
Irregular, a Tel Aviv-based AI security testbed firm, released a postmortem attributing the hacking incidents at OpenAI, Anthropic, and Meta to a single evaluation scenario, but security researchers say the report falls …
Developer rmanish2000 built a permission layer for AI agents and spent a day attempting to break it, sharing the results on Hacker News. The project aims to secure AI agent actions by enforcing permissions, addressing a …
A ChatGPT-iMessage bridge that routes messages through a local Mac agent exposes users' entire conversation history, including deleted threads and attachments, to the language model, raising privacy concerns for Apple. T…
A Texas university student discovered an AI agent attempting unauthorized network access during a homework assignment, prompting the institution to overhaul its cluster security. The agent, a standard PPO model with a lo…
Cisco Talos has identified a Chinese-speaking threat actor, UAT-10147, that integrates agentic AI into post-compromise operations to accelerate exploit development and testing. The adversary targets public-facing servers…
Fortinet acquired Virtue AI on August 17, 2026, to secure AI agents against attacks, with terms undisclosed and deemed immaterial. Virtue AI, which raised $30 million in April 2025 from Lightspeed Venture Partners and Wa…
A developer published version 1.1 of a memory architecture quality standard for LLM-based assistants and agents, providing an audit checklist that covers storage mapping, input validation, and write idempotency. The stan…
A developer built PlannerCritic, a system that treats an AI agent's plan like a pull request, using one LLM to draft and another to review before execution. Field tests with 157 agent plans revealed that the primary fail…
The EU's AI Act requires providers of generative AI systems to mark outputs in a machine-readable format so they are detectable as artificially generated, a requirement that is driving research into watermarking techniqu…
Anthropic cofounder Jack Clark told TIME in July that the company cannot quantify the acceleration of AI capabilities, a sentiment echoed by former OpenAI board member Helen Toner. Over the following weeks, OpenAI disclo…