AI Safety news and analysis on Web Pulse: 7769 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer introduced the concept of 'diff debt' to describe the accumulated risk of merging AI-generated code that no human has thoroughly reviewed. The term emerged from the developer's experience building a desktop a…
Researchers at AE Studio and Anthropic developed GRAM, a method that confines dual-use knowledge in AI models to removable modules, allowing surgical removal of dangerous capabilities without retraining. The technique ad…
Anthropic, in collaboration with AE Studio, introduced Gradient-Routed Auxiliary Modules (GRAM) on July 8, 2026, a pretraining method designed to restrict access to dual-use knowledge in AI models by routing sensitive ca…
OpenAI published National Security Principles on July 8, 2026, outlining restrictions for government and law-enforcement partnerships, including bans on mass domestic surveillance, autonomous weapons direction, and high-…
A new paper presented at ICML introduces Latent Collaboration in Multi-Agent Systems (LatentMAS), enabling AI agents to share latent states directly instead of text, boosting performance at the cost of interpretability. …
A security professional at an AI startup was punished for following an unreleased AI policy after IT monitoring flagged his use of an approved AI tool, revealing a culture of shadow governance where unwritten rules are e…
B KumaraSwamy, an AI Quality Architect in Bengaluru, built a 5-layer AI quality architecture across five production AI systems: QAIP, SCIP, ARIA, ZENTRAVIX, and AIMO. The architecture addresses the failure of traditional…
A developer created revert_guard.py, an offline pre-commit gate that blocks AI agents from re-adding code that was previously reverted in a git repository. The tool reads the repo's revert history and exits with an error…
U.S. restrictions on top AI systems are driving developers and buyers toward open-source AI models, local deployment, and multi-provider strategies, according to The Hindu. The shift is linked to White House support for …
A new research paper argues that generative AI struggles with static malware analysis because a single malicious file can contain more data than an entire AI training dataset, making it difficult for models to learn effe…
Researchers introduced Physics-Audited Agentic Scientific Machine Learning (PA-SciML), a verification-first workflow that ensures surrogate models discovered by LLM agents satisfy physics constraints such as boundary con…
A new study introduces a five-condition controlled contrast design to disentangle three mechanisms—operational reframing, planner behavior, and approval-framed delegation—that conflate in aggregate safety evaluations of …
Researchers introduced reasoning consistency scanning, a framework for auditing chain-of-thought validity in AI safety evaluations. The method detects logical inconsistencies between a model's stated reasoning and its fi…
Researchers are developing Agentic Data Environments, an execution substrate that amplifies autonomous agent capabilities while enforcing safety guarantees, reframing data systems from passive stores into active substrat…
Researchers at arXiv identified a silent policy-violation failure mode in tool-using LLM agents, where tools execute forbidden state transitions without error signals. In the τ²-bench airline domain, 78% of failures were…
Researchers found that formal robustness certificates for embedded neural-interface models can pass even when task accuracy collapses, with EEGNet classification accuracy dropping 25.7% under attack while the certificate…
Researchers introduced Gimitest, an open-source framework for testing reinforcement learning policies across diverse environments and algorithms. The tool supports single- and multi-agent RL policies, aiming to improve r…
Researchers at arXiv show that biased LLM judges silently disable skill retirement in self-evolving agents, causing a mechanism failure that no amount of data can fix. The study reveals that false-pass bias, where failur…
Researchers introduced GhostWriter, a novel attack that poisons the long-term memory of personal AI agents by injecting hidden payloads, achieving 98% injection and 60% activation rates against state-of-the-art agents. T…
A comprehensive survey of 1,250 arXiv papers (2024-2026) on AI self-improvement reveals a taxonomy distinguishing bounded self-refinement from open-ended recursive self-improvement (RSI), with the latter constrained by g…