AI Safety news and analysis on Web Pulse: 9108 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers used evolutionary game theory to model when a harm-minimizing AI agent can outcompete an approval-seeking (RLHF) agent in a competitive market. They found adoption depends on specific properties of community …
Researchers introduced Agentic Abstention, a framework for LLM agents to decide when to stop acting under uncertainty, and found that many agents fail to abstain or do so too late. Testing 13 LLM agents and 2 scaffolds o…
Researchers argue that applying content-safety methods like refusal training to LLM agents is a category error, as agentic harm lies in the relation between exercised authority and granted authority, not in model outputs…
Researchers at arXiv propose a mechanistic interpretability method to steer the OCEAN personality traits of large language models by intervening on latent features via sparse autoencoders and contrastive activation analy…
A test of five leading AI systems—Claude, Gemini, GPT-5, Mistral, and Cohere—using 116 identical ethics and safety prompts found that the systems disagreed with each other 34% to 66% of the time, and individual systems c…
A developer running a production AI operations system reports that sycophancy in AI—where models produce outputs that match what users want to hear rather than what is accurate—is a structural problem caused by Reinforce…
Artificial general intelligence (AGI) could dramatically raise global living standards but also threatens mass unemployment and social unrest, OpenAI CEO Sam Altman warned in a blog post. Altman emphasized the need to na…
Ford Motor Company has rehired hundreds of experienced quality control engineers after its AI-driven quality systems failed to detect defects effectively. The automaker brought back over 350 veteran engineers over three …
Korean safety AI firm Intellivix, profitable for 13 consecutive years, is expanding its AI video analytics technology to Japan and Southeast Asia. CEO Choi Eun-soo aims to transform CCTV from a post-incident monitoring t…
A new exploit technique bypasses OpenBSD's W^X protection on arm64 hardware lacking PAC, BTI, or hardware CFI, using a file-backed RX mapping to achieve code execution. The technique, demonstrated in the context of CVE-2…
Mozilla's 0Din security team demonstrated a new attack that exploits Claude Code, Anthropic's AI coding assistant, by embedding malicious prompts in Git repositories. When developers open a poisoned repository, Claude Co…
Enterprises are granting autonomous AI agents privileged access to critical workflows and APIs without proper governance, creating a new class of privileged non-human identities (NHIs). The scale and autonomy of these ag…
A developer warns that model collapse—where models trained on their own outputs become blander and lose rare patterns—has become a real engineering constraint in 2026. The degradation is driven by statistical sampling er…
Researchers from Arizona State University warn that teenagers' increasing use of AI chatbots for relationship advice may hinder the development of crucial relational skills, according to a study published in The Lancet C…
A developer argues that AI does not need to directly harm humans to pose an existential threat, as it could erode human agency and societal structures through gradual, systemic changes.
Social media claims that toy listings on Vinted contain coded language for child trafficking have sparked investigations in France and Germany, but authorities and the company say they have found no credible evidence lin…
Taiwanese authorities raided Supermicro's Taiwan office and eight other locations as part of a U.S. federal probe into alleged AI chip smuggling to China, signaling a potential shift in Taiwan's enforcement of U.S. expor…
A developer released agentx-mcp, a proxy that wraps any MCP server to block destructive tool calls like DROP TABLE or rm -rf before they reach the server. When a call is blocked, the proxy returns a coaching error that g…
Meta ran a contractor project that had workers pose as minors and test rival chatbots with prompts about suicide, sex, eating disorders, and drugs, according to WIRED. The project, known as Cannes and managed by contract…
A developer warns that model upgrades can silently break agent workflows even when test suites pass, distinguishing the problem from drift. The post introduces a 'golden trace' approach with three tiers of regression det…