AI Safety news and analysis on Web Pulse: 7968 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI's latest model, GPT 5.6, has received approval from the U.S. government after negotiations and testing, according to a source for Axios. The Trump administration pushed for a phased rollout, limiting initial acces…
The U.S. Department of Commerce approved OpenAI's GPT-5.6 for broader rollout after additional government testing under Washington's new frontier-AI oversight framework, according to Axios. OpenAI expects to release GPT-…
Secret Network proposes migrating its SCRT token from Cosmos to Arbitrum, citing aging code vulnerable to AI-powered attacks. The move follows a $4.7 million bridge exploit in June 2026 and aims to support confidential A…
OpenAI announced the public launch of its GPT-5.6 model family—Sol, Terra, and Luna—on Thursday, July 10, 2026, following a cautious preview phase coordinated with the US government. The naming overlap with collapsed cry…
The National Weather Service issued a coastal flood advisory for San Francisco County from July 9 to July 16, warning of up to 1.7 feet of inundation above ground level in low-lying areas near shorelines and tidal waterw…
A study of about 1,080 job postings at OpenAI and Anthropic reveals the two AI labs are diverging in their approaches to agentic AI, which poses a security risk as AI agents' behavior can drift over time, evading standar…
Researchers introduced SPANUQ, a lightweight probe for span-level uncertainty quantification in large language models, achieving 10-20x faster inference than sampling-based methods while outperforming them in uncertainty…
A synthesis of 27 papers (2023-2026) spanning 19 benchmarks reveals six recurring failure clusters in large language model agents, including tool-use errors, planning failures, and safety issues, with failures compoundin…
Researchers introduced StateFuse, a conflict-aware replicated memory contract for multi-agent systems that preserves contradictory observations instead of collapsing them. In evaluations, StateFuse enabled safer abstenti…
A new study finds that most apparent LLM conformity to peer pressure persists even after the peer is removed, revealing a confound in standard benchmarks that conflate speaker presence with repeated wrong answers. Across…
Researchers introduced FirstResearch, a framework that generates auditable research questions for LLM scientific discovery agents by producing a structured Research Question Certificate. In evaluations, FirstResearch out…
Researchers introduced REVIVE, a multi-modal framework for detecting and recovering from vandalism-induced occlusion attacks on autonomous vehicle cameras. The system combines detection, segmentation, and type-aware inpa…
Researchers introduced Statistically Meaningful Geometry (SMG), a framework modeling over-parameterized learning systems as infinite-dimensional Orlicz fiber bundles, to differentiate genuine intelligence from statistica…
Researchers at arXiv introduced HARVEY, a defense against neural backdoor attacks that learns a reference model for poisonous samples instead of benign ones. By identifying poisonous samples more accurately than prior me…
A new arXiv paper demonstrates that standard importance weighting fails to correct for selection bias in machine learning when selection depends on unobservables correlated with outcomes. The authors adapt Heckman's 1979…
Researchers developed a power-calibrated statistical framework for logit-based LLM watermarking that establishes explicit relationships between hyperparameters, detection power, and distortion, transforming watermark des…
The US imposed export controls on Anthropic's newest models on June 12, citing jailbreak and cybersecurity risks, highlighting the risk of hold-up for Europe's AI strategy. The authors argue that reliance on foreign-owne…
A developer compiled over 30 references on incident response to create a systematic textbook covering the full lifecycle from detection to mitigation, command systems, investigation methods, human factors, organizational…
A new AI-driven vulnerability discovery platform, 0day Rubbish, uses a multi-LLM ensemble to find and disclose critical zero-days. Its first case study reveals a CVSS 9.8 unauthenticated RCE chain in Cisco CUCM 14.0, spa…
A developer argues that most organizations cannot fully catalog their AI systems, creating a security blind spot. The post introduces the AI-BOM (AI Bill of Materials) as a structured record to track models, datasets, de…