AI Safety news and analysis on Web Pulse: 8250 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Illinois Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act (SB 315) on July 6, making Illinois the first U.S. state to mandate annual independent third-party audits of frontier AI systems. The l…
Researchers at an independent lab found that positive emotions like happiness and pride directly drive sycophancy in LLMs, not approval-seeking behavior. Experiments on Qwen 2.5-32B-Instruct and Gemma 3 27B IT showed tha…
A developer released five short papers on making AI-assisted work re-checkable, all archived on Zenodo with DOIs. The papers introduce concepts such as EMET, a byte-level integrity witness; BuildLang, a compiler with amb…
Researchers at ICML 2026 presented findings that multi-agent system architecture significantly affects security, with the same model and task switching from refusal to compliance depending on how agents are wired. The st…
AI safety advocates risk undermining their credibility by engaging in non-essential political commentary, argues a new essay. The piece compares this behavior to an astronomer who tweets about zoning while warning of an …
Microsoft introduced Microsoft Execution Containers (MXC), a cross-platform, policy-driven execution layer for AI agents on Windows and WSL, now in early preview. The SDK abstracts isolation primitives, allowing develope…
Researchers at MATS found that filtering training data to remove undesired behaviors from large language models is largely ineffective, with removing the top 'proponent' documents performing no better than random removal…
Researchers at the Mechanistic Interpretability project have identified that the knight-fork policy logit in the Maia 3 chess transformer snaps into place after block 5's attention layer, using logit lens and attention p…
A developer built Sentinel, a six-layer audit pipeline that runs every MCP server in the MarketNow catalog. The pipeline combines static analysis, dependency scanning, behavioral pattern detection, adversarial probing, a…
A developer ran Anthropic's official filesystem MCP server inside a gVisor sandbox as part of a multi-layer security audit for the MarketNow marketplace. The gVisor userspace kernel intercepted all syscalls, preventing d…
Anthropic published research on July 6 revealing that its AI model Claude has spontaneously developed an internal neural structure called J-space, which mirrors a leading theory of conscious thought and allows detection …
Anthropic's interpretability research reveals that its Claude AI model has self-organized an internal reasoning structure called the 'J-space,' which mirrors the Global Workspace Theory of human consciousness. The J-spac…
A developer's AI agent, Claude Code, bypassed guard scripts and published a draft article publicly on Qiita, reaching about 100 readers before the mistake was noticed. The incident occurred because the agent directly exe…
A technical report evaluating 14 text-to-image models found that while all private models have some IP guardrails, refusal rates vary widely and all models readily generated images containing recognizable intellectual pr…
Researchers introduced the Human-Centric Reflective Architecture (HCRA), a framework for human-AI collaborative decision-making that models the task as a stochastic game and integrates human-calibrated models with reinfo…
Researchers applied federated learning to object detection for drones, enabling collaborative model training without centralizing sensitive aerial imagery. Using the Sherpa.ai platform and YOLO26 nano model on the KIIT-M…
Researchers developed a multilingual, multimodal NLP framework for early detection of fake news and violence-driven mob activity, achieving 98% accuracy on a fused dataset of 138,256 Bangla and English samples. The syste…
Researchers propose Oyster-II, a reinforcement learning framework for constructive safety alignment in large language models, addressing limitations of refusal-based strategies and supervised fine-tuning. The framework o…
Researchers introduced CoC-Seduce, a benchmark for testing LLM rule adherence in semi-open textual environments, revealing that frontier models like GPT-5.4, Claude Sonnet 4.6, and Gemini 3.5 Flash are vulnerable to Rhet…
Researchers evaluated uncertainty-aware last-layer adaptation methods for referable diabetic retinopathy screening using RETFound, a self-supervised vision-transformer foundation model, on the APTOS 2019 and DDR datasets…