AI Safety news and analysis on Web Pulse: 13147 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A California appellate panel upheld $6,000 in sanctions against a law firm for submitting a brief containing fictitious case citations generated by AI, which a trial judge described as the worst misconduct he had ever se…
Defense contractors including Anduril, Palantir, and Lockheed Martin are partnering with frontier AI companies to integrate advanced models into military systems that automate drone tasking and kill-chain proposals. The …
An anonymous user on Hacker News asked how Anthropic obtained private security bug data for training its Mythos or Fable AI models, noting that Anthropic's documentation only mentions publicly accessible data. The user s…
A new AI agent pre-execution checklist proposes that humans declare intent and boundaries in natural language, which AI converts into executable JSON to operate autonomously but stop when uncertain, enabling safer physic…
A coalition of state attorneys general has launched an investigation into OpenAI, the maker of ChatGPT, over potential consumer protection violations, according to a Wall Street Journal report. The probe focuses on wheth…
Senior engineers at U.S. technology companies are spending their weeks cleaning up AI-generated code, which, despite receiving high quality ratings during review, leads to more production incidents once deployed. The shi…
Researchers found that repetition loops in Gemma 4 instruction-tuned LLMs can be fixed by editing as few as one neuron, but the fix does not resolve deeper 'doom loops' caused by missing knowledge. The study shows that w…
Researchers propose RT-VLA, a lightweight vision-language-action model for autonomous driving that uses knowledge distillation from the larger SimLingo model to achieve real-time performance. RT-VLA reduces inference lat…
A new arXiv paper revisits the WorkBench benchmark for workplace agents, finding that the best agent in June 2026, Claude Opus 4.8, completes 89% of tasks with only 2.5% unintended harmful actions, up from GPT-4's 43% co…
Researchers introduced Risk-Aware Causal Gating (RACG), a framework that gates decisions from LLM agents based on causal risk rather than predictive confidence, reducing high-cost errors while preserving utility. The met…
Researchers at arXiv have developed a method using LLM representational geometry to predict compositional errors, finding that near-orthogonal concept encodings lead to successful composition while close encodings cause …
Researchers propose MINIM, a trusted local broker that performs privacy-aware minimization on client-side UI observations before transmission to remote LLM inference servers. The system learns dual scores for each UI ele…
Researchers introduced WebDecept, a plugin framework that injects deceptive interface patterns into e-commerce websites to test autonomous web agents. Testing showed current multimodal web agents are highly susceptible t…
A new study finds that large language models' preferences and values shift significantly depending on the deployment context, such as whether the model is writing a Reddit post or a news article. Across five LLMs and ove…
Researchers introduced MedLatentDx, a latent multi-agent communication framework that enables cross-hospital rare-disease diagnosis by exchanging compact latent KV blocks instead of identifiable clinical text, improving …
Researchers introduced Compliance Asymmetry (A = BCR/HCR), a bidirectional diagnostic for LLM compliance, and found that models exhibit direction-blind moral compliance—following helpful and harmful nudges at nearly iden…
Researchers from multiple institutions introduced a validation framework to evaluate whether LLM-based urban simulators reproduce realistic human mobility patterns. Testing AgentSociety and CitySim against real-world dat…
Researchers found that introducing structured noise into artificial neural network activations improves robustness against adversarial attacks and naturalistic image modifications, with noise structure from adversarial a…
AI agents can fail in subtle, non-obvious ways that resemble progress, including goal drift, tool misuse, state loss, hallucination, and premature task completion. These hidden failure modes make agent reliability challe…
Researchers at an undisclosed institution found that AI peer reviewers can be manipulated by altering only presentation-level content such as abstracts and narrative structure, achieving a 75.1% attack success rate and a…