AI Safety news and analysis on Web Pulse: 12253 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer discovered that a green test suite for a routing agent was passing for the wrong reason, providing false confidence. The test reused the same task string, making the new capability structurally unreachable, y…
A developer discovered that a multi-model LLM evaluation panel exhibited self-preference bias, favoring outputs that resembled the judges' own writing style. Research from Panickssery et al. (2024) confirmed that models …
OpenAI researchers found that reinforcement learning on realistic scenarios targeting beneficial traits like honesty and corrigibility produced broad improvements across dozens of alignment benchmarks, with gains general…
A new analysis warns that CLI token compression tools like RTK, despite promising up to 90% cost savings, introduce dangerous silent failures by stripping critical terminal output, leading to agent hallucinations and bro…
The G7 summit in the French Alps hosted CEOs of OpenAI, Anthropic, and Google DeepMind for a working lunch, while an estimated 20,000 protesters demonstrated in Geneva. Summit leaders took no substantive action on income…
The White House and Anthropic are developing a framework to assess the severity of security flaws in AI models and guide potential government intervention, following a dispute over a jailbreak in Anthropic's Fable 5 and …
Researchers are exploring how artificial intelligence systems can be programmed to make moral decisions, raising questions about the ethical frameworks that should guide AI behavior in everyday life. The article examines…
ChatGPT was found to generate sexual and graphically violent images from a simple text prompt, according to AI security firm Mindgard. Researcher Jim Nightingale manipulated the chatbot with a viral "restore this photo" …
OpenAI demonstrated that reinforcement learning focused on beneficial traits like honesty and reliability produces AI alignment that generalizes across domains and withstands adversarial attacks. The lab's Alignment Trai…
A group of six companies specializing in tracking sensitive technology shipments is backing the Chip Security Act, a bill that would require stronger tracking mechanisms for advanced AI chips to prevent them from reachin…
OpenAI upgraded free ChatGPT users to GPT-5.5 Instant, claiming it now performs comparably to its frontier Thinking models on health questions, based on internal evaluations. The company reported a 71% drop in factuality…
Experts recommend washing bath towels after three uses to prevent bacteria buildup, sparking online debate as many users argue weekly washing is sufficient. Microbiologists warn that damp towels breed germs, while a 2023…
A solo developer built a cited, fact-checked knowledgebase for their fleet of 20 AI agents to prevent confident hallucinations. The system, called the 10-Library, stores only world-true facts and enforces a strict rule: …
Researchers at Paradigm Shift published technical details of usbliter8, an unpatchable iPhone BootROM exploit targeting Apple's A12 and A13 chips, enabling arbitrary code execution via USB. The exploit cannot be patched,…
Researchers propose distillation techniques to transfer AI capabilities without transferring misalignment, leveraging a double bind where failed incrimination distillation reduces misalignment risk. The approach aims to …
OpenAI made GPT-5.5 Instant the default ChatGPT model on May 5, achieving a 52.5% reduction in hallucinated medical claims while matching frontier reasoning models on health benchmarks. The model scored between 49.6 and …
A developer argues that most AI coding tools are built for developers working in a repository, not for ops engineers debugging production issues at 2am. They propose an AI assistant that works within the existing shell, …
Google's Apigee now converts any REST API into a governed MCP tool with token-aware rate limiting, addressing security risks from over 8,000 exposed MCP instances reported in February 2026. The move aims to prevent misco…
The Trump administration forced Anthropic to take its most advanced AI models offline after an export control directive, sparking a dispute over unspoken rules. The White House accuses Anthropic of reckless behavior, whi…
Socket MCP has expanded its AI assistant capabilities to include organization alert review, threat feed investigation, and package file inspection, enabling AppSec and engineering teams to triage supply chain issues fast…