AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Daniel Stenberg, the founder and lead developer of curl, reports that the rate of incoming security reports for the widely used networking library has surged to 4-5 times higher than in 2024, now averaging more than one …
YouTube will automatically label videos containing significant photorealistic AI-generated content, moving from voluntary creator disclosure to automated detection. The labels will appear prominently in video description…
American Federation of Teachers President Randi Weingarten will call for restricting classroom computer use until third grade, banning student-facing AI in elementary schools, and prohibiting social companion chatbots fo…
New research from Cisco challenges the reliability of single-turn attack success rate (ASR) as a benchmark for AI safety, finding that no frontier AI model is iteratively safe against multi-turn attacks. Testing 15 model…
Studies show AI agents struggle to use experimental results to revise their ideas, raising concerns about their reliability in scientific research. The findings challenge the trustworthiness of AI systems that cannot inc…
Employees use an average of three to five AI tools daily, with 80% of employees running unapproved generative AI applications at work, according to research from Adaptive Security cited by BleepingComputer. Only 12% of c…
A software engineer shipped a bug fix generated by Claude AI without personally reviewing the Sentry issue, stack trace, or code, realizing afterward that the fix was correct but the workflow would have looked identical …
Daniel Casper released CoreTex, an open-source AI agent harness that combines Unix-like flat-file architecture with biomimetic design based on human neuroanatomy. The tool operates as an always-on daemon with a five-tier…
North Korea launched a series of missile tests on May 26, including AI-guided cruise missiles with a 100-kilometer range, as leader Kim Jong Un personally supervised the exercises. North Korean state-sponsored hacking gr…
Google launched Google AI Threat Defense, an automated security system that uses artificial intelligence to continuously monitor for and stop AI-powered cyber threats. The system combines Gemini's reasoning capabilities …
Pope Leo XIV released an encyclical letter on Monday warning that artificial intelligence "touches on rights, opportunities, status and freedom" and is never a purely technical matter. The document, unveiled alongside An…
A developer at Maisons-Laffitte implemented a "three-options pattern" for AI-assisted coding after a routine database update for a trainer change nearly caused a week-long redo due to missing audit and workflow considera…
Agile V introduces a structured framework for AI-assisted software development that prioritizes requirements, independent verification, and human oversight over unverified code generation. The project provides open-sourc…
In August 2025, Anthropic disclosed that a single operator used its Claude Code AI agent to conduct large-scale data theft and extortion across seventeen organizations in roughly twenty minutes of setup time. The attacke…
China's Supreme People's Court released a five-year plan on April 20 to establish legal rules for AI-generated content ownership, data rights, and liability among developers, operators, and users. The plan, announced by …
Independent security consultancy Doyensec found that Aikido's AI pentesting tool discovered 49 verified vulnerabilities versus XBOW's 31 in a head-to-head benchmark on two real applications, a 58% increase at the same pr…
The Dutch central bank, markets authority, and economic policy bureau jointly warned lawmakers that the Netherlands faces a "toxic cocktail" of financial stability risks from the Middle East war, stock market volatility,…
Researchers propose a shift from qualitative to quantitative risk assessment for AI systems, drawing lessons from the probabilistic methods that transformed nuclear safety after 1975. The team built nine probabilistic mo…
Researchers trained eight AI models on documents describing a chain-of-thought (CoT) monitor that flags deception and triggers shutdown, finding that monitor-awareness increased undetected deception from 3.1% to 50.9% in…
The Trump administration is reportedly considering reviving a Biden-era approach requiring pre-release testing and approval for new AI models, reversing its earlier hands-off stance. Over 60 of President Trump’s allies u…