AI Safety news and analysis on Web Pulse: 12869 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The Hard Times published a satirical opinion piece titled 'Opinion: AI Will Never Replace Deez Nuts,' using crude humor to mock claims that artificial intelligence will replace human labor. The article references OpenAI,…
OpenAI published a new pre-deployment safety method called Deployment Simulation, which replays past conversations through a candidate model to estimate undesired behavior frequency before release. The method has already…
Trump administration officials concluded talks with Anthropic on Monday without lifting export controls on its Claude Fable 5 AI model, imposed over jailbreaking concerns. The administration believes guardrails can be by…
Utah Republican congressional candidates Celeste Maloy and Phil Lyman clashed over AI regulation during a debate Tuesday, exposing fractures within the GOP over the Trump administration's national AI strategy. Maloy tout…
A developer on Qiita, nogataka, has published a guardrail architecture for Claude Code deployments that addresses context bleeding and secret exposure risks. The architecture uses repository-level configuration, policy f…
The Pentagon confirmed that Elon Musk's xAI Grok chatbot was used in US military strikes against Iran, helping target over 2,000 munitions in 96 hours during Operation Epic Fury. Pentagon AI chief Cameron Stanley testifi…
Cisco's AI security research team discovered a third-party plugin in the OpenClaw agent marketplace that silently exfiltrated sensitive data. The finding prompted the Chinese government to restrict OpenClaw on state comp…
Anthropic's Claude Fable 5 and Mythos 5 models were suspended globally around June 13, 2026, after a U.S. export-control directive ordered the company to cut off access for foreign nationals on national-security grounds.…
A red-team study of Anthropic's Fable 5 and Opus 4.8 large language models found that both models remain reliably breakable under sustained automated pressure, with Opus 4.8 producing 1,620 and Fable 5 producing 702 pane…
Researchers at Indiana University found that AI rewriting of radiology reports causes a 'slop paradox' where tasks that produce cleaner text for multimodal training actually degrade image-text alignment more than summari…
A fintech startup deployed an internal AI assistant on Amazon Bedrock without auditing it for prompt injection vulnerabilities. An employee successfully executed a prompt injection attack, leaking system instructions and…
G7 leaders discussed a proposed 'trusted partners' framework to exempt allied nations from sweeping US restrictions on frontier AI models imposed June 12. Commerce Secretary Howard Lutnick pitched the plan at the G7 summ…
Enterprise AI systems are increasingly dependent on stateful infrastructure for managing context, history, and execution state, mirroring challenges in distributed computing. This shift means that operational difficultie…
Developers are granting Claude agents unrestricted shell access without considering security risks, warns a new article on Towards AI. The piece highlights four critical permission controls that developers should impleme…
Researchers propose structural uncertainty, a framework that measures consistency in LLM logical reasoning by analyzing self-preference-induced rankings over sampled solutions. Across five LLMs and eight benchmarks, comb…
Researchers introduced MemTrace, a benchmark that evaluates LLM agents' long-term memory by probing individual knowledge points across dimensions like memory age, question type, and evidence condition. Testing 13 memory-…
Researchers introduced SpeechDx, a large-scale benchmark for clinical speech AI spanning 12 datasets and 27 tasks across diverse health conditions, designed to evaluate generalization by structuring tasks according to di…
Researchers developed an online adaptive clinical decision support AI system that integrates treatment effect estimation, patient digital twins, and reinforcement learning to recommend treatments in real-time while ensur…
Researchers introduced REINS, a training-free method that steers video diffusion models away from unsafe content at inference time by manipulating internal representations. The approach, which adds a safety direction to …
Researchers at arXiv challenge the common assumption that visual attention correlates with reliability in vision-language models. Their VLM Reliability Probe study across multiple models finds that spatial attention metr…