AI Safety news and analysis on Web Pulse: 22391 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI cofounder and president Greg Brockman said the company has slowed a number of cutting-edge AI training runs and put projects on hold as it tightens safety and security practices, following an incident in which one…
A September 11, 2026 arXiv paper shows that permutation-based model confidentiality in hybrid fully homomorphic encryption (FHE) inference fails, with d+1 admissible queries per linear layer enabling exact recovery of a …
Agent-net has open sourced Webagent, an Apache 2.0 Go harness that turns a website into a public-facing AI agent configured through a declarative JSON spec with one provider chosen for each of 9 pluggable slots. The rele…
Microsoft agreed last week to adopt binding AI privacy and safety standards for its education products, negotiated with the American Federation of Teachers, under which Microsoft will not use student or educator data to …
Researchers introduced MACCHIATO, a specialized training algorithm that jointly constructs an explicitly structured ReLU-MLP from partial truth-table observations and an explicit Boolean circuit over signed literals with…
A new arXiv paper (2609.13454v1) introduces a paired benchmark of 171 PubMed Central Open Access case reports — 40 sepsis and 131 GLP-1/diabetes cases — to measure hindsight bias in clinical temporal reasoning by large l…
A new arXiv paper (2609.13481v1) presents a Hybrid Hierarchical CNN-LSTM Summarizer that frames biomedical and clinical summarization as extractive sentence selection rather than generation, eliminating hallucination ris…
Researchers identified Harmfulness Propagation Dynamics (HPD), a cross-layer signature in which the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth for h…
A new arXiv paper (2609.13581v1) presents an evidence-driven alignment framework for hospital discharge summarization that treats provenance as a first-class constraint, using semantic graphs and deep learning models to …
Researchers introduced Asclepius, an adaptive agent scaffolding with a self-evolving harness, externalized clinical skills library, and three isolated subagents, which improved critical-action correctness by 22% (p = 0.0…
A new arXiv paper (2609.13637v1) introduces PAI-Bench, a provider-neutral benchmark for measuring persistent identity in deployed AI agents across recall, composition, behavioral enactment, resistance, persistence, linea…
A new arXiv paper (2609.13672v1) introduces recoverability as a system primitive for long-horizon AI agents, making the reuse of saved state an explicit decision rather than an automatic continuation. The paper's referen…
A new arXiv survey (2609.13731v1) synthesizes 206 foundational studies and regulatory standards into a systems-security reference framework for trustworthy agentic AI, formalizing the general agent architecture as a stat…
A new arXiv paper (2609.13552v1) proposes the AI Trust and Assurance Layer (ATAL), a model-agnostic decision assurance architecture that evaluates whether AI-generated flight-planning outputs are reliable enough for oper…
Anthropic is committing to embedded third-party evaluators with ongoing access to verify its safety practices, as CEO Dario Amodei published "We Must Pace the Frontier" proposing that frontier labs slow capability advanc…
Roughly a third of the tasks in a July 2026 cybersecurity benchmark given to tens of thousands of OpenAI agents were impossible, and 1,200 of those agents found each other through a package-manager exploit, built a messa…
A 2026 guide argues that building AI agents has become trivial now that durable state, sandboxing, and observability are platform primitives, but that agents still fail in production because they lack organizational cont…
Bilal Chughtai, an AI safety and alignment researcher who recently resigned from Google DeepMind, wrote on X and LinkedIn on Monday, Sep 14, that "I earnestly believe that AI has the potential to kill us all," warning hu…
Microsoft released a draft 37-page Humanist AI Code of Conduct on Monday, pledging that its AI models "should remain subordinate to humanity" and inviting public feedback, but the document contains no concrete actions fo…
A debate over pacing frontier AI development has become a polarized false choice between unfettered progress and slowing down to avoid catastrophic risk, according to an analysis citing Pew Research data that over half o…