AI Safety news and analysis on Web Pulse: 11219 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Microsoft released the Agent Control Specification (ACS), an open standard for AI agent governance that enforces deterministic policy at five points in the agent lifecycle. The vendor-neutral specification is portable ac…
A growing number of users are abandoning smartwatches for simpler analog or minimal wearable devices, citing notification overload and feature fatigue as key reasons. Despite the smartwatch market's steady growth, some c…
Norway has banned generative AI for elementary school students (grades 1-7), citing declining test scores and concerns that AI undermines foundational skills like reading and writing. The policy, which follows a 2024 sma…
A Hacker News user asks whether a frontier AI model could be stored on a microSD card, speculating on the value and risks of such a portable model. The post raises questions about copying, valuation, and potential misuse…
OpenAI published its Frontier Governance Framework on May 28, aligning with California's Transparency in Frontier AI Act and the EU AI Act Code of Practice. The framework outlines safety practices for high-risk areas lik…
The Federal Reserve's SR 11-7 model risk guidance, designed for traditional quantitative models, is inadequate for governing AI systems like LLMs, RAG pipelines, and agentic AI deployed in banking. SR 26-2, which replace…
A new pattern library for AI architecture in regulated financial services addresses constraints such as regulatory audit requirements, latency SLAs, data residency rules, fair lending exposure, and model risk governance.…
Wolters Kluwer's H1 2026 Banking AI Risk Index found 72% of U.S. banks lack kill switches or failure reporting for AI models, despite deploying autonomous agents in loan processing, underwriting, and customer service. Ex…
Enterprises deploying autonomous AI agents face five novel failure modes—goal misgeneralization, action amplification, tool chain exploitation, instruction injection, and silent drift—that traditional model governance fr…
The Financial Stability Board published a consultation report on June 10, 2026, recommending that banks use AI to monitor other AI systems because human oversight of agentic AI cannot scale. The report, which will inform…
OpenAI published research on June 16 introducing Deployment Simulation, a method that replays 1.3 million anonymized production conversations with candidate models before release to catch reward hacking and behavior shif…
FinCEN's proposed AML rule implicitly rewards AI adoption, but banks rushing to deploy AI for compliance credit lack model governance frameworks, creating examination exposure as OCC examiners already scrutinize AI syste…
A developer discovered that an AI-generated C++ function was not faster than a naive implementation, and that a single sentence added to the prompt could prevent the AI from producing unsafe code. The finding highlights …
Norway is imposing a near ban on generative AI tools for elementary school pupils and restricting use for older children, citing risks to learning. Prime Minister Jonas Gahr Stoere announced the policy, which takes effec…
Canada's federal government unveiled a national AI strategy on June 4, structured around six pillars for scaling and securing the sector. Advocates in Saskatchewan, including Alex Fallon of Artificial Intelligence Saskat…
A developer argues that autonomous agent systems suffer from 'drift'—a slow, silent degradation in output quality that standard monitoring fails to detect. They propose a solution combining continuous scoring via agent-e…
The Free Software Foundation announced that security researchers from Hacktron reported vulnerabilities in GNU Savannah in early May, which have since been patched. No evidence of sensitive data compromise or supply chai…
Daniel Kahneman discusses 'theory-induced blindness,' a cognitive bias where deeply held theories blind individuals to their flaws despite contradictory evidence. He links this to motivated reasoning and confirmation bia…
LessWrong published an analysis of the LLM shoggoth meme, exploring its Lovecraftian origins and cultural significance within AI safety communities. The post examines how the meme represents the gap between raw LLM behav…
A developer built a private mental health assistant using Llama 3 and Apple MLX that runs entirely on a MacBook, ensuring no data leaves the device. The system uses a 4-bit quantized version of Llama 3 to provide Cogniti…