AI Safety news and analysis on Web Pulse: 12873 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers introduced SpeechDx, a large-scale benchmark for clinical speech AI spanning 12 datasets and 27 tasks across diverse health conditions, designed to evaluate generalization by structuring tasks according to di…
Researchers developed an online adaptive clinical decision support AI system that integrates treatment effect estimation, patient digital twins, and reinforcement learning to recommend treatments in real-time while ensur…
Researchers introduced REINS, a training-free method that steers video diffusion models away from unsafe content at inference time by manipulating internal representations. The approach, which adds a safety direction to …
Researchers at arXiv challenge the common assumption that visual attention correlates with reliability in vision-language models. Their VLM Reliability Probe study across multiple models finds that spatial attention metr…
Researchers propose RepSelect, a method for robust LLM unlearning that isolates forget-set-specific representations by collapsing top principal components of weight gradients, achieving 4-50x larger reduction in post-rel…
Researchers propose MODE-RAG, a multi-agent system using Variational Free Energy and attention states to dynamically gate interventions in multimodal retrieval-augmented generation, reducing hallucinations and logical fa…
Researchers introduced STATEWITNESS, an activation explainer that decodes hidden states from reasoning LLMs to audit deceptive behavior, achieving 0.916 mean AUROC and outperforming existing monitors by 11.6% to 25.0%. T…
Researchers at the University of Toronto introduced a novel method to evaluate second-order social bias in large language models (LLMs), where models exhibit bias in their judgments about biased content. Using a reasonin…
Researchers formalized and verified four concurrency anomalies in multi-agent LLM systems, establishing a machine-checked consistency hierarchy from L0 to L4. They implemented verified detectors and runtimes in Rust, pre…
Researchers at arXiv have identified a conflict signature in language models that distinguishes deceptive outputs from honest errors, achieving 100% accuracy in detecting lies across multiple models including GPT-2, Qwen…
The United States has decided not to blacklist Chinese AI company DeepSeek, while designating over 100 other Chinese firms as national security risks, signaling a targeted approach to technology restrictions.
Nvidia CEO Jensen Huang said in an Associated Press interview that society must adopt new social norms as artificial intelligence becomes more widely used, urging people to engage with AI. He supported government regulat…
OpenAI researchers tested whether public chat data from WildChat can predict real-world AI misalignments, finding that deployment simulations using public conversations can estimate rates of undesirable model behavior, o…
KPMG, a global consulting firm with 250,000 employees, published an AI-generated report in October 2025 titled 'Total Experience: Redefining Excellence in the Age of Agentic AI.' An audit by GPTZero found that 40 of 45 c…
The US government pressured AI company Anthropic to restrict access to its latest models, 'Fable 5' and 'Mythos 5', to US citizens only, highlighting Europe's digital sovereignty challenges ahead of the G7 summit.
Tide's team of five engineers and designers built a verifiable authentication system that eliminates the need for a central credential store, removing a common breach vector. The team includes cryptography experts, a Key…
QuantaMind built a headless CLI tool to shift AI evaluation left into CI/CD pipelines, allowing teams to gate model merges based on agentic reliability thresholds. The tool integrates custom eval JSON collections into th…
An attacker hijacked a Mastra maintainer's account and republished 116 packages in the @mastra catalog over 27 minutes, adding a hidden dependency on the typosquat package easy-day-js. The malicious package disables TLS …
South Korea became the fourth country to form an AI security alliance with OpenAI, signing an MOU to collaborate on global AI safety evaluation frameworks and develop assessment standards reflecting Korean language and s…
The US Commerce Department has delayed adding Chinese AI developer DeepSeek, memory chipmaker CXMT, and over 100 other Chinese tech firms to the Entity List for nearly eight months, the longest gap in over a decade. The …