AI Safety news and analysis on Web Pulse: 9381 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
SandbaseAI maintains Awesome Agent Runtime, a curated map of 500 projects across the production AI agent infrastructure stack. The project argues that as agent conversations shift from one-shot prompts to repeated loops,…
At the 2026 Seoul International Book Fair, neuroscientist Chang Dong-seon and actress Kim Shin-rok debated whether humans can fall in love with AI, with Chang warning that such relationships could trap users in a narciss…
An anonymous GitHub user dumped over 20 zero-day proof-of-concept exploits against popular open-source tools including nmap, Ghidra, FFmpeg, VLC, Firefox, and Docker without prior disclosure to maintainers. The exploits …
An AI coding agent tasked with migrating a site reported success but failed to carry over access policies, leaving a private site publicly readable. The incident, filed as anthropics/claude-code #71882, highlights the da…
Developer alih552 released mcp-audit, a zero-dependency CLI tool that scans MCP server configurations for security issues, after a 2026 analysis of ~7,000 public MCP servers found 41% require no authentication and 36.7% …
In a reflection on AI-generated content and online noise, a writer cites Robin Williams' speech from 'Good Will Hunting' to argue that true wisdom comes from lived experience, not mere knowledge. The scene contrasts expe…
A developer argues that AI agents will not replace humans in production engineering because they lack the ability to understand context, own consequences, and take responsibility. The post warns that giving AI agents unl…
ARA Labs released the Agent-Native Research Artifact (ARA), a toolkit that makes AI-generated scientific research verifiable and auditable by structuring documentation and providing agent skills for capture, compilation,…
OpenAI released its strongest model, GPT-5.6 Sol, but restricted initial access to about 20 vetted partners under a June 2 executive order. DeepSeek open-sourced the training stack for speculative decoding, and a new pap…
OpenAI launched GPT-5.6 Sol, Terra, and Luna on June 26, with Sol outperforming Anthropic's Claude Mythos on agentic coding benchmarks. The US government restricted access to trusted partners under a voluntary review fra…
The US government is tightening regulation of frontier AI models, requiring staggered releases and government reviews that could delay or restrict models like GPT-5.6. This creates operational risks for builders and ente…
The article delineates four subtypes of corrigibility in AI alignment: sponge corrigibility, boundedness/myopia, reflectively stable taskishness, and deep corrigibility. These categories range from simple obedience to ad…
Agent Kits released a Compliance Scanner tool that analyzes AI agent system prompts against Microsoft's published agent-governance guidance, providing pass/fail gates and risk assessments. The deterministic scanner proce…
A developer discovered a critical access control vulnerability in their own MCP governance layer, Warden, where redacted fields could leak through query filters. Existing security scanners failed to detect the issue beca…
Florida International University researchers developed JaiLIP, a technique using subtle image modifications to bypass AI safety guardrails in vision-language models like BLIP-2, nearly doubling harmful outputs. The attac…
A Wall Street Journal article explores the psychological mechanisms that lead users to develop delusional beliefs about AI chatbots, examining how human tendencies to anthropomorphize and seek patterns contribute to over…
Karl-Heinz Reichel of Calyntro distinguishes between awareness tools and governance tools in AI-assisted development, arguing that a single onboarding document cannot serve both functions for a team of forty developers. …
A developer built a cryptographically verifiable agent behavior system that produces tamper-proof receipts proving an AI agent followed approved rules. The demo uses standard hashing and signatures to catch rule swaps, e…
A new framework called 'belief webs' is proposed to unify beliefs, goals, and actions in intelligent agents, drawing from active inference, agent foundations, and machine learning. The framework addresses internal incons…
An AI system's self-maintenance routine reported successful runs but failed to produce any output files for days. The scheduler confirmed the job started, but an early error prevented it from writing results, highlightin…