AI Safety news and analysis on Web Pulse: 22277 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A study of 22 large language models found that closed-source frontier systems — including GPT-4.0-5.2, Grok-3/4, Gemini 2.5 Pro/Flash, and Claude Sonnet 4.5/4.6 — produce self-reported personality profiles that align wit…
A study applying the RIFT global rubric failure taxonomy to two clinical benchmarks found that an LLM judge flagged 29.6% of HealthBench Professional criteria as non-atomic and 65.4% as misaligned or rigid. Rewriting bun…
Researchers proposed RAG-CT, a defense that detects malicious queries in Retrieval-Augmented Generation (RAG) systems by analyzing their entropy and margin distributions with a score-based detection method, according to …
A new arXiv paper (2609.16056v1) formalizes behavioral knowledge as a precondition Bayesian network over an agent's structural actions and tests three injection points in a neuro-symbolic RL loop: a symbolic verifier use…
A position paper posted to arXiv (2609.16189v1) argues that no language-model-enabled wargame should inform planning, doctrine, policy, or crisis response without an auditable safety case, and that open-ended wargames sh…
A new arXiv review (2609.16213v1) finds that AI uplift for biological threats currently exists primarily in digital tasks, with frontier systems exceeding expert baselines on in-silico and screening-evasion benchmarks, w…
A narrative review published on arXiv identifies eight recurring ethical and privacy issues in LLM-enabled geospatial AI (GeoAI), including passive location inference from mobility traces, spatially structured bias ampli…
A new arXiv paper (2609.16245v1) introduces Metacognitive Steering, an inference-time controller that reads the cognitive regime of Kimi 2.6, a trillion-parameter mixture-of-experts model, and dynamically composes layer-…
Researchers proposed CLEAR, an agentic framework for cross-source evidence adjudication in large language models in medicine, detailed in arXiv paper 2609.16301v1. CLEAR independently generates candidate answers from thr…
Researchers introduced RAIL, an interface exposing polyhedral enclosures of nonlinear dynamics to LiRPA-style bound propagation, and CLIPPER, a branch-and-bound algorithm that jointly refines enclosures and splits contro…
Researchers introduced Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using LLM agents, containing 22 attack families and 35 scenarios across seven domains that yield more than 2,500 …
A new arXiv paper (2609.16592v1) introduces an end-to-end framework for generating context-specific large language model benchmark datasets by combining expert input with synthetic data generation. The framework uses a s…
A new arXiv paper (2609.16589v1) reports that large language models possess an intrinsic value system, based on projecting responses from 106 LLMs at 150,000 queries per model and 95,000 human survey profiles into a shar…
A new arXiv paper (2609.16730v1) introduces LSREP, a Longitudinal State-Replay Evaluation Protocol for conversational memory, with ICE v2 as its audited local-first architecture case study. The private single-user instan…
A new arXiv paper (2609.16779v1) presents the first end-to-end approach that enables large language models to perform the complete Analytic Hierarchy Process (AHP) workflow, according to its authors. The study constructs…
US Senator Chris Murphy (D-CT) said in a four-minute video posted to X on Wednesday 16 September that the AI industry "can't" protect Americans, warning that leaving the job to the industry means "we are cooked, as an ec…
A developer has described an AI software pipeline built around four sequential agent roles — a writer, checker, arbiter, and final gate — that exchange only fixed PASS/FAIL signals rather than free-text handoffs, with an…
Meta CEO Mark Zuckerberg publicly rejected calls for a coordinated slowdown in AI development on September 16, 2026, arguing that market competition and legal liability already give labs sufficient incentive to build saf…
House Speaker Mike Johnson said Congress should not impose a federal moratorium on artificial intelligence development, arguing a pause would cost the United States its edge over China and carry "serious national securit…
Anthropic CEO said he "didn't appreciate" the speed with which AI's growth would underpin the global economy, in remarks published 16 Sep 2026. His call to slow down AI development sent shockwaves through global stock ma…