AI Safety news and analysis on Web Pulse: 8253 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers propose Oyster-II, a reinforcement learning framework for constructive safety alignment in large language models, addressing limitations of refusal-based strategies and supervised fine-tuning. The framework o…
Researchers introduced CoC-Seduce, a benchmark for testing LLM rule adherence in semi-open textual environments, revealing that frontier models like GPT-5.4, Claude Sonnet 4.6, and Gemini 3.5 Flash are vulnerable to Rhet…
Researchers evaluated uncertainty-aware last-layer adaptation methods for referable diabetic retinopathy screening using RETFound, a self-supervised vision-transformer foundation model, on the APTOS 2019 and DDR datasets…
Researchers identified five failure modes in perturbation-based construct-validity audits of AI safety benchmarks, demonstrating that implementation details can silently invalidate audit conclusions. In a case study of t…
Researchers propose Lagrangian Reward Augmentation (LARA), a framework for inference-time alignment of language models that enforces safety constraints by dualizing a constrained optimization problem into a one-dimension…
A July 5, 2026 arXiv paper proposes a Generative AI Control Framework to help financial institutions align generative AI governance with the Federal Reserve's SR 26-2, which modernizes model risk management. The framewor…
By 2026, AI agents moved from demos to production, introducing runaway costs from tool calls and recursive reasoning. A VentureBeat survey found 79% of enterprises took financial hits from autonomous agents, while Uber b…
Researchers introduced CAGE-1, an evaluation framework for enterprise agentic AI that assesses control, assurance, and governance before deployment. The framework tests authority, policy enforcement, retrieval quality, m…
Researchers propose MPSelectTune, a two-stage fine-tuning method that improves concept unlearning in large language models by focusing on the prompt type with the highest concept accuracy. The approach achieves 2-15% mai…
Researchers introduced SecureCROWN, the first framework for privacy-preserving neural network robustness verification using secure two-party computation. It allows a model owner and data owner to jointly compute certifie…
Researchers propose single-prover interactive proofs for AI safety, eliminating the need for debate between two AI models. The new method works for oracle-aided computations under robustness or low-degree polynomial assu…
Researchers have developed a method to distinguish AI-generated fake speech from natural human speech by analyzing vowel spectral distributions using the Wasserstein metric and persistent homology, with Japanese as a tes…
Hugging Face introduced a new Kernels update on July 6, adding a dedicated Hub repository type, trusted publisher gating, code signing via Sigstore, and enhanced provenance metadata to make custom AI kernels a safer, mor…
Highloop LLC's Strata app delivers AI-powered backcountry route planning with a plain-English safety verdict of go, watch, or avoid, drawing on avalanche bulletins, slope data, and live weather via Anthropic's Claude. Th…
Anthropic CEO Dario Amodei revealed in a Bloomberg interview that he has only one direct report, chief of staff Avital Balwit, while his sister and co-founder Daniela Amodei oversees all other executives as president. Th…
Panoptes, an AI audit and alignment layer, launches to provide universal observability, audit trails, and policy enforcement for AI coding agents like Claude Code, OpenAI Codex, Google Antigravity, and Hermes Agent. The …
China's new AI regulations take effect July 15, 2026, forcing ByteDance's Doubao and Alibaba's Qwen to disable agent and persona features. The 'Interim Measures for the Administration of AI Anthropomorphic Interactive Se…
YouTuber Husk IRL placed three phones running different conversational AI agents together and asked them to count to 100. The agents negotiated a shared approach but repeatedly agreed with and reinforced each other, fail…
AI workloads are driving uncontrolled cloud costs, with some environments seeing expenses double in 30 days due to ungoverned LLM endpoints, vector databases, and GPU instances. The problem mirrors earlier cloud sprawl b…
Sysdig's threat research team caught what appears to be the first ransomware attack run almost entirely by a large language model, named JADEPUFFER. The AI agent broke into a server, stole credentials, encrypted a databa…