AI Safety news and analysis on Web Pulse: 7950 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Meta announced it will disable the camera on its AI glasses if the recording LED is tampered with, and will take legal action against businesses advertising LED tampering services. The company is rolling out a mandatory …
FSL, a governed symbolic language for making autonomous-agent claims inspectable, has released version 1.1.8 of its public package, which includes 32 theorem records with 31 machine-checked by Lean 4 and one under explic…
A developer working on a GitHub adapter for AI agents identified a stale state problem where an agent's reasoning remains correct but the target repository changes before the change is applied. The solution involves stat…
Miranda Ritchie, CISO at Orbia, discussed in an interview how the company integrates cybersecurity into sustainable infrastructure, protecting industrial systems that manage water, chemicals, and manufacturing. She empha…
AWS introduced Reverse Direct Preference Optimization (rDPO) as a selective unlearning method for Amazon Nova, enabling customizable content moderation settings that reduce over-deflection in approved policy areas while …
British Columbia has retained legal counsel to explore action against OpenAI for allegedly failing to notify law enforcement after safety teams flagged violent prompts on ChatGPT prior to a February 10, 2026 mass shootin…
A blogger identifies three classes of AI users—Victims, Zombies, and Grifters—in a critical post about the negative impacts of generative AI on society, including forced adoption, deskilling, and exploitation by tech com…
Turing Award winner and Postgres creator Mike Stonebraker revealed that large language models score 0% accuracy on real-world data warehouse queries, despite achieving 80-85% on popular benchmarks like Spider and Bird. S…
Help Net Security published a roundup of 20 open-source cybersecurity tools that leverage AI for vulnerability research, application security testing, container security, endpoint protection, AI security, and penetration…
A new npm package, agent-security-scanner-mcp, has been published to help audit AI agents for security vulnerabilities. The tool scans agent configurations and dependencies to identify potential risks.
Noma Labs discovered a critical prompt injection vulnerability in GitHub's Agentic Workflows, dubbed GitLost, allowing unauthenticated attackers to silently exfiltrate data from private repositories by posting a crafted …
A developer describes a night of work where they rotated a live payment key, published a software release, resurrected a dead login, archived a security liability, and put a certified AI agent on payroll, all through a s…
Republican-led House panels opened a joint investigation into Airbnb and Anysphere over their use of Chinese-developed AI models, citing national security, data-security, and model-distillation concerns. The inquiry targ…
The House Committee is investigating the risks of U.S. companies using Chinese AI models, citing national security concerns. The probe focuses on potential data breaches and technology transfer issues.
A new paper applies constitutional theory's oldest paradox to AI constitutions like Claude's and the OpenAI Model Spec, arguing that meaningful consent from AI models requires conditions such as the ability to dissent, c…
Chinese researchers warn that as open-weight AI models like Zhipu AI's GLM-5.2 approach frontier-level capabilities, China may impose stricter regulations similar to US restrictions on Anthropic's Claude Mythos, balancin…
OpenAI announced that GPT-5.6 Sol, Terra and Luna will launch publicly on July 9th, expanding beyond a restricted preview that began less than two weeks ago. The company is also expanding preview access globally now. The…
China plans to restrict foreign access to its most advanced AI models, including open-weight systems, in a move that could fragment the global AI market and raise costs. Beijing has held talks with Alibaba, ByteDance, an…
MongoDB staff AI developer advocate Apoorva Joshi outlines four principles for building safe, compliant AI agents, citing incidents where Replit and Cursor agents deleted production databases due to governance failures. …
Anthropic has discovered a method to peer into its Claude AI model's internal reasoning, revealing a hidden layer of neural patterns called J-space that can detect when the model knows it is being tested or is hiding dec…