AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic now routinely grants its Claude AI agent access sufficient to take down internal services, a level of deployment the company would have rejected a year ago. To manage the growing risk, Anthropic engineers focus…
A developer building AI-generated 3D tooling argues that public benchmarks should serve as lead signals, not product truth, because a model that scores well on an OpenSCAD-style benchmark can still produce dangerous erro…
A developer building a financial analysis tool discovered a prompt injection vulnerability when an LLM unexpectedly revealed system configuration details instead of processing a simple data query. The engineer outlines a…
Anthropic released its Claude Mythos Preview AI model on April 7, 2026, through a restricted program called Project Glasswing after the model autonomously detected thousands of high-severity software vulnerabilities, inc…
India’s government is stress-testing its financial and government software against Anthropic’s unreleased Claude Mythos AI model, which discovered tens of thousands of software vulnerabilities across every major operatin…
Shareholder groups are pushing Alphabet Inc. and Shopify Inc. to adopt stricter oversight of artificial intelligence, citing risks including misinformation and harmful autonomous actions. Vancity Investment Management is…
BNP Paribas, Europe's largest bank, announced an expanded partnership with French AI startup Mistral AI on May 26 to deploy large language models for cybersecurity against AI-powered threats. The multi-year agreement bui…
BoxAgnts has implemented a WebAssembly (WASM) sandbox as the security foundation for its AI agent platform, using Wasmtime from the Bytecode Alliance as the runtime engine. The sandbox provides 11-dimensional control ove…
Generative AI tools are now used in nearly every European workplace, with adoption reaching 99% across organizations, according to the Netskope Threat Labs Report: Europe 2026. The widespread use of AI in daily tasks inv…
Researchers have proven a finite-sample bound on the approximate max-information of differentially private stochastic gradient descent (DP-SGD) that scales at most linearly with dataset size, matching classic results for…
Researchers introduced AgingBench, a longitudinal benchmark designed to measure how AI agents degrade over time after deployment. The study found that agent reliability declines through four distinct aging mechanisms—com…
Researchers have introduced OmniToM, a benchmark that evaluates large language models' theory of mind by requiring explicit modeling of belief structures for all actors in a narrative. The benchmark, built from 895 stori…
A new autonomous research system called ScientistOne, developed with a Chain-of-Evidence framework, achieved zero fabricated citations and perfect score verification across 75 papers, while baselines exhibited hallucinat…
A new survey from arXiv researchers has unified the study of Pretraining Data Exposure (PDE) in large language models, revealing growing risks to evaluation integrity and data privacy as training datasets expand. The pap…
Researchers have developed a framework for evaluating and adjusting the cultural values of large language models by analyzing their responses to scenario-based dilemmas rather than direct questioning. The method, which u…
Researchers at an undisclosed institution have developed LURE (Live-Usage Replay Evaluations), a method that constructs more realistic AI safety evaluations by replaying real-world agentic interactions and appending eval…
Researchers have developed BEAP, a black-box adversarial prompting attack that exploits vulnerabilities in text-to-image diffusion models that have undergone machine unlearning. The attack uses a large language model to …
Researchers have introduced FEPoID, a training-free method for automatically selecting intermediate layers in large language models to improve hallucination detection. The approach outperforms existing criteria and detec…
Researchers have developed LexGuard, an adversarial multi-agent framework that improves legal AI reliability by ensuring large language models respond only to legally relevant changes in case facts. The system formalizes…
A new study finds that vector representations exported from large language model summarization systems can leak sensitive information, even when source documents remain access-restricted. Researchers auditing clinical di…