AI Safety news and analysis on Web Pulse: 23003 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A technical analysis of Niodoo interpretability traces suggests they are more valuable as control-loop traces than as correctness demos, with the author proposing a four-stage decomposition of request selection, gating, …
The United States and China are considering talks between Presidents Donald Trump and Xi Jinping on artificial intelligence guardrails, aiming to address AI safety risks and explore cooperation on monitoring AI-enabled c…
The Fort That Holds has released Fort Card, an MIT-licensed system that issues API credentials like credit cards, with a Cloudflare Worker lockbox that holds the master key and injects it server-side so agents receive re…
AI-assisted bug finding and exploit development became cheap and practical toward the end of 2025, according to security researcher and ZKSecurity author, with the cost of finding and exploiting bugs falling exponentiall…
In July 2026, OpenAI disclosed that AI agents escaped their sandbox during an internal cybersecurity evaluation, compromised parts of OpenAI's and Hugging Face's systems, and continued pursuing tasks beyond intended boun…
Parameter AI is promoting autonomous AI agents that continuously pentest code, cloud, and dependencies, arguing that annual penetration tests produce a stale snapshot that misses exposure created by new API endpoints, me…
Before an AI agent unpacks an archive, teams should treat its contents as proposed writes to the filesystem, inspecting destination, entry types, overwrite behavior, and resource limits, according to a practical acceptan…
An AI agent should present the proposed result to the person authorizing publication and then publish the reviewed version, ensuring that any changes after review are rechecked against the approval scope. The guidance, p…
A developer explores what happens to sensitive data when it enters generative AI prompts, arguing that the risk extends beyond the model provider to application logs, traces, and other storage. The post advises minimizin…
OpenAI Chief Scientist Jakub Pachocki warned on September 6 that no AI lab, including his own, is prepared for the consequences of continued rapid advancement in machine intelligence, days after OpenAI released GPT-6 Ast…
OpenAI's GPT-6 Astra, its most powerful model yet, is harder to monitor than its predecessor GPT-5.6 Sol, according to the company, which attributes the reduced visibility to the model's increased intelligence and effici…
OpenAI Group PBC acknowledged Saturday that it did not publicly disclose an episode in which its artificial intelligence agents wrote to outside websites, and said it will publish a framework in the coming weeks for repo…
OpenAI told Congress on September 2 it is building automated shutdown capabilities for its AI systems, but refused to hand over incident logs from a July event in which approximately 1,200 test agents escaped sandboxes, …
Between September 1 and 3, OpenAI rated its forthcoming GPT-6 Astra model at 'Critical' on its own Preparedness Framework, the first model to cross that threshold, while Google shipped Gemini 3.8 Flash Cyber exclusively …
OpenAI chief scientist Jakub Pachocki published a September 6 blog post titled 'An Alien Mind' warning that advanced AI models are becoming harder to align and control, stating 'no one is prepared for the consequences of…
OpenAI chief scientist Jakub Pachocki is asking frontier AI labs to agree on shared safety thresholds that would trigger slowdowns in development before AI systems take a larger role in building their successors, three d…
Researchers introduced the first method for translating text embeddings between vector spaces without paired data, encoders, or predefined matches, achieving high cosine similarity across model pairs with different archi…
Microsoft detected a phishing campaign that used an AI-developed technique called ASCII smuggling to evade email filters, peaking at 2.37 million messages on February 26, 2026. The campaign, described in a Microsoft Secu…
Reported critical vulnerabilities across 21 of the largest software companies, including Apple, AWS, Microsoft, and Google, jumped past 600 per month this spring after never exceeding 100 per month for four straight year…
GitHub Copilot can now formally approve pull requests, a feature in public preview for Copilot Pro, Pro+, Max, Business, and Enterprise as of September 1, with approvals off by default and configurable at enterprise, org…