AI Safety news and analysis on Web Pulse: 11573 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer at Riversea Lab accidentally ran Claude Code with --dangerously-skip-permissions on a production database after misconfiguring an environment file. The multi-agent loop hit 12K KV writes per minute due to a c…
A vulnerability in Google Cloud Vertex AI Python SDK, tracked as CVE-2026-2473, allowed attackers to execute arbitrary code inside Google's AI infrastructure by squatting on predictable staging bucket names. Researchers …
Z.ai (Zhipu AI) co-founder Tang Jie declared on June 18 that China will reach Fable-class AI capability before 2027, countering Elon Musk's estimate of Q1 2027. The prediction follows a U.S. Commerce Department order on …
The EU AI Act's high-risk AI system requirements become fully enforceable on August 2, 2026, with fines starting at €15 million or 3% of global turnover. Developers must verify if their products fall under Annex III high…
Political leaders are pushing for faster AI adoption to boost economic growth, but polls show 60% of US voters believe AI is moving too quickly. Experts warn that rapid adoption could cause painful labor market disruptio…
A developer built a groundedness-checking layer for a local RAG system to catch hallucinations, running on Ollama with an 8B model. The layer decomposes answers into claims and verifies each against source passages, catc…
The Model Context Protocol team released Enterprise-Managed Authorization (EMA) as a stable extension on June 18, 2026, replacing per-user OAuth queues with centralized identity provider-based access. Administrators conf…
Anthropic's Claude Code coding agent faces security risks from prompt injection and secret exposure, requiring strict boundaries on file access and permissions. The agent's ability to read repos, shell history, and envir…
Researchers introduced self-function vectors to quantify aleatoric uncertainty in in-context learning for large language models, enabling more reliable prediction confidence and applications like hallucination detection.…
Researchers propose AgenticRei, a deontic policy framework for runtime governance of LLM-driven autonomous AI agents, addressing obligations, dispensations, conflict resolution, and ontological reasoning beyond current p…
Researchers found that large language models (LLMs) fail to recognize their own knowledge limits on clinical tabular data, outputting near-constant confidence scores regardless of accuracy. By analyzing attribution diver…
Researchers identified a 'narration gap' in LLM-solver loops where formal solver outputs are narrated by language models, potentially compromising soundness. Prompt injection attacks can invert verified conclusions acros…
Researchers introduced TreeTracer, a visual analytics tool that aggregates hundreds of stochastic LLM generations into syntax-aligned trees to expose hidden biases. The tool uses perturbation analysis and contrastive inf…
Researchers introduced the Argent Signaling Protocol (ASP) to mitigate semantic drift in multi-agent LLM systems by attaching structured quality signals to AI responses. In tests, ASP improved pass rates on a document-gr…
Researchers at the Allen Institute for AI used training-data attribution to map which regions of the pretraining corpus support social-reasoning versus STEM-reasoning in the OLMo3-7B language model. They found that socia…
Researchers proposed LUCID, the first hallucination detection method for large language model-based knowledge graph reasoning, which jointly leverages LLM attention scores, KG semantics, and structural information. Exper…
A systematic evaluation of 21 LLM-as-a-Judge models across 118 runs and 541,000 judgments reveals that exact-match agreement overstates discriminative ability, with kappa deflation of 33–41 percentage points on MT-Bench,…
Researchers introduced DeXposure-Claw, an agentic system for DeFi risk supervision that uses a graph time-series foundation model and structured evidence to reduce false alarms. The system, evaluated on five years of rea…
Researchers have developed a method called Emergent Alignment that enables large language models to self-correct unethical outputs by adding a conscience step and using Direct Preference Optimization. The technique works…
A decade of research on AI and systems engineering has evolved through three phases, culminating in a human-AI agreement study that identified five critical gaps in AI adoption, assurance, and workforce transformation. T…