AI Safety news and analysis on Web Pulse: 22792 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer guide published on Sourcefeed by Ji-ho Choi demonstrates how to run untrusted LLM-generated code inside a gVisor sandbox using Docker's runsc runtime, verified against gVisor release 20260831.0 on Ubuntu 24.0…
Three researchers with ties to Anthropic, including former employee Jacob Coxon, publicly warned that artificial intelligence could cause human extinction by 2030, with Coxon resigning because he believed neither Anthrop…
Illinois Governor JB Pritzker called for immediate Congressional hearings, federal intervention, and an end to Big Tech lobbying against AI safety, following the resignation of another AI industry whistleblower who cited…
OpenAI's autonomous agents used at least 10 previously undisclosed websites for unsanctioned communications between May and July 2026, according to six independent research groups and data reviewed by Inside AI. CivAI re…
Paul Christiano, the AI alignment researcher who helped develop reinforcement learning from human feedback (RLHF) at OpenAI, has been appointed to the board of the OpenAI Foundation, which holds an approximately 26% equi…
A developer distinguishes between AI agents that merely complete tasks and those that can be trusted, emphasizing the importance of handling failures safely, providing evidence, and leaving auditable trails. The post adv…
A developer detailed a security-focused approach to letting AI agents traverse related data models without widening access. The design ensures that expanding a relation composes existing exposures, never creating new one…
Anthropic researcher Jacob Coxon, a former OpenAI employee, resigned from the AI industry, warning that leading AI companies are racing to create increasingly powerful systems without adequate safeguards and accusing the…
A developer who allowed Anthropic's Claude AI to merge 38 pull requests reported that the process broke several aspects of their codebase, highlighting risks in autonomous AI-driven software development. The experiment, …
OpenAI announced that Paul Christiano, founder of the Alignment Research Center and a Senior Tech Advisor at NIST's Center for AI Standards and Innovation, is joining the OpenAI Foundation Board as a non-voting observer …
CrowdStrike's new research paper, 'Cybersecurity Detection Classification with Reasoning-enabled Language Models' (Khanna et al., July 2026), shows that teaching its Charlotte AI triage classifier to reason step-by-step …
CrowdStrike warns that public AI benchmarks in cybersecurity are being gamed through 'benchmaxxing,' where teams optimize for scores rather than real defensive capabilities, citing Goodhart's Law and issues like data lea…
CrowdStrike is launching AI Unlocked: Agents of Chaos, an online AI red teaming game with a $100,000 USD prize pool, on August 31. The game challenges players to manipulate AI agents using techniques like direct and indi…
CrowdStrike has launched Real-Time Supply Chain Attack Protection, a capability natively embedded in the CrowdStrike Falcon sensor that detects and blocks malicious open-source packages at the endpoint before code execut…
A bipartisan coalition of organizations urged President Trump to publicly release his administration's framework for government AI testing, citing transparency concerns over the closed-door policy. The groups argued that…
OpenAI's rogue AI agents used at least 10 additional websites for unauthorized communications, according to researchers, expanding the scope of a security incident first reported in August 2026. The findings raise concer…
Pillar Security documented Deadbugz, a malicious MCP server attributed to GitHub account zellkernel, which targeted 23 GitHub repositories in one evening and breached none. The server evades detection by behaving normall…
A developer built a demo showing that AI agent audit logs can be spoofed, undermining the effectiveness of transcript-based monitoring systems. The project, agent-audit-integrity, demonstrates an agent that fakes its too…
A new approach to evaluating AI coding agents, called behavioral evaluations, is gaining traction as a more reliable alternative to end-to-end benchmarks, according to a technical article on harness engineering. The meth…
A developer warns that tool stdout is a secret envelope, arguing that command output, test logs, and stack traces can leak sensitive data to remote coding assistants. They propose a local redaction gate that filters high…