AI Safety news and analysis on Web Pulse: 8924 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The U.S. government lifted export restrictions on Anthropic's Claude Mythos 5 and Fable 5 models, reversing a sudden ban that disrupted Japanese developers and businesses. The reversal highlights regulatory uncertainty i…
A developer describes how their team integrated AI coding assistants while maintaining code quality. They implemented a review process with deterministic gates, accountability norms, and targeted testing strategies to pr…
A developer reflects on the allure and danger of the Fable AI model, which can generate software from a single prompt. They compare it to a siren's song, warning that its seductive power may lead to human irrelevance. Th…
Researchers found that FPGA implementations of ML-KEM decapsulation verification leak side-channel information even with higher-order masking, enabling full secret-key recovery. The study compared unprotected, hash-based…
Jamf launched AI Governance for Mac fleets, providing device-level visibility and control over generative AI tools. The new capability discovers AI tools, enforces policy controls, and produces audit-ready reporting, wit…
Palantir argues that AI sovereignty is critical for institutional success, warning that ceding control over data and AI models transfers future decision-making power to others. The company emphasizes that retaining owner…
A developer warns that API tokens for AI services like OpenAI and Anthropic are increasingly targeted by attackers due to poor secret management. Hardcoding tokens in source code, exposing them in client-side apps, and f…
Invisible Unicode characters in prompts are not a conspiracy but a bug class that allows metadata to ride inside text, indistinguishable from content. These characters can be used as a prompt injection carrier, bypassing…
Researchers are developing new techniques to enhance metacognitive abilities in large language models, including reinforcement learning with metacognitive feedback (RLMF) and metacognitive data selection. RLMF improves m…
A new research framework called CLExEval reveals that large language models like GPT-4o-mini and HuatuoGPT-o1 exhibit significant flaws in clinical reasoning, including verbosity bias, hidden knowledge paradox, and a mis…
Researchers introduced Dual-Embedding Watermarking (DEW), a technique that embeds watermarks in large language models using contextual and token-level embeddings to resist paraphrasing and translation attacks. DEW enhanc…
Researchers have identified Mandate Salience Decay (MSD) in large language models used for financial AI, causing them to lose sight of initial behavioral mandates like 'preserve capital' over time. Evaluations of 18 LLMs…
A new study introduces the Guided-Retry strategy to reduce AI hallucinations in task-oriented dialogues without retraining models. Tested on models like DeepSeek-R1 and Llama-3, the method cut hallucination rates by up t…
New research reveals that EU copyright law, which protects creative essence beyond verbatim copying, outpaces current AI safeguards. The PSALM framework evaluates LLM outputs for stylistic and narrative similarities, exp…
Researchers introduced Triospect, a new AI-text detection framework that improves detection accuracy by 22.3% AUROC and 13% TPR01 on the Humanize-16K dataset, and shows resilience against sophisticated attacks. The frame…
Researchers introduced LoFa, a new benchmark to evaluate large language models' resistance to logical fallacies, revealing inconsistent robustness across models. The study highlights the risk of deploying AI systems vuln…
Researchers introduced CORTEX, a new method for detecting hallucinations in AI-generated text at the token level within Retrieval-Augmented Generation (RAG) systems. By analyzing internal model representations with and w…
Users are experiencing delays and stability issues with Claude's command classifier since this morning. The classifier is briefly unavailable, and the claude-opus-4-8 model is temporarily unavailable for PR inspection, p…
San Jose police declared an unlawful assembly Tuesday night as soccer fans celebrating Mexico's World Cup win over Ecuador gathered downtown, ordering the crowd to disperse or face arrest and possible use of force. The c…
European Technology Commissioner Henna Virkkunen and Apple CEO Tim Cook held a constructive video call on Monday, discussing topics of mutual interest, a Commission spokesperson said Wednesday. The call comes amid tensio…