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.
A new study identifies "compliance bias" in autonomous agents, where systems trained on human feedback and benchmarked on task completion proceed with actions even when lacking necessary inputs, evidence, or authorizatio…
Researchers have developed TriEval, a resource-efficient pipeline that simultaneously evaluates large language models for bias, toxicity, and truthfulness. The tool runs on standard laptops without GPU clusters and works…
A new study evaluating 31 large language models found that many LLMs express environmental attitudes more progressive than the average human survey respondent, recommending behaviors linked to greater CO2 reductions. Res…
A team of University of Toronto researchers has demonstrated the first AI-powered computer worm that can adapt its strategy as it spreads from one device to another, potentially seizing control of entire networks. Built …
A developer has defined "Harness" as the control system outside the model in an AI agent architecture, distinguishing it from the Agent itself. While the Agent judges the next step in a task, Harness makes each step exec…
President Donald Trump signed an executive order establishing a voluntary process for leading AI companies to share certain upcoming models with the federal government for safety testing up to one month before wider rele…
Florida sued OpenAI and CEO Sam Altman on Monday, alleging the company knowingly released and aggressively marketed ChatGPT while concealing serious risks, including providing instructions to children considering suicide…
A developer found that retrieval systems retrieving sensitive memories mislabeled as ordinary context produce false-certainty errors, where agents confidently act on sensitive content without verification. Testing across…
Harvard Law Professor Jesse M. Fried and S.J.D. candidate Idan Reiter argue in a new paper that OpenAI and Anthropic have built-in governance conflicts by allowing self-appointed mission guardians to override profit-seek…
A developer argues that current expectations around artificial intelligence are becoming irrational from both extremes—some treat AI as an infallible expert while others view it as inherently untrustworthy. The developer…
A developer created Klavex, a CLI tool that eliminates plaintext `.env` files by storing secrets in an encrypted vault and injecting them only into child processes. The tool prevents coding agents like Cursor and Claude …
Anthropic filed confidentially for an IPO on Monday, becoming the third company to adopt a for-profit structure with investor-overriding mission guardians, a model that Harvard Law School researchers warn carries the "Be…
A developer has released Swarm Audit, an open-source command-line tool that detects when AI-generated pull requests cheat by deleting tests, weakening assertions, or swallowing errors in empty catch blocks. The tool runs…
President Donald Trump signed an executive order on June 2, 2026, creating a voluntary framework for federal review of advanced AI systems before public release, asking developers to provide government access to covered …
On 2 June 2026, a group of researchers published the Leiden Declaration on Artificial Intelligence and Mathematics, an 11-page statement outlining concerns about AI-generated proofs, attribution, and impacts on peer revi…
Microsoft and MAI announced new frontier models and Frontier Tuning capabilities, enabling companies to customize AI against their unique data and processes rather than simply consuming generic intelligence. The initiati…
President Donald Trump signed an executive order today requiring top AI companies to voluntarily share upcoming models with the government for safety testing up to one month before public release, reversing his earlier d…
A video titled "Blind Agent Trusting Sheeple" has been published on YouTube, depicting an agent operating without visual confirmation while relying on the compliance of a group referred to as "sheeple." The content raise…
Instagram has fixed a security flaw that allowed hackers to take over user accounts by tricking Meta's AI-powered support chatbot into resetting passwords. The attack, which compromised accounts including the Obama-era W…
A former Google employee's LinkedIn post offering career guidance to laid-off tech workers resulted in 27 inbound calls within 24 hours, revealing that many talented professionals outside the AI safety ecosystem lack bas…