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 senior data engineer at a multinational bank discovered a $12 million batch settlement error caused by a stale configuration table, while an LLM-driven customer-support assistant suggested a loan product violating inte…
Gary Bernhardt argues that the control layer above AI models, not the models themselves, is the true product in agentic AI systems. He advocates for deterministic state machines, typed boundaries, and guardrails to ensur…
Since May 28, ChatGPT has been prepending an undocumented "hidden user memory" check phrase to some responses without user or developer notice. OpenAI has issued no changelog or documentation for the behavior, which comm…
AI coding tools have become so essential to developers that a majority now refuse to work without them, even for research purposes, according to a February 2026 study by AI research lab METR. However, mounting evidence s…
AI-assisted development, often called "vibe coding," accelerates prototyping but creates a significant gap between functional demos and production-ready software, according to an engineer's analysis. Senior engineers mus…
A lone attacker published 14 malicious npm packages that mimicked popular OpenSearch and Elasticsearch libraries. Microsoft identified and removed the packages, which were designed to trick developers into downloading ma…
Okta has developed a new tool that allows customers to instantly terminate rogue AI agents, addressing growing enterprise concerns about uncontrolled autonomous software. CEO Todd McKinnon confirmed that clients includin…
A developer built BiasLens, a free tool that scans job rejection letters for documented discrimination patterns under US employment law in 30 seconds. The hardest engineering challenge was forcing the AI to accurately re…
Anthropic released Claude Opus 4.8, an incremental upgrade to its AI model, just six weeks after Opus 4.7. The new model is smarter, can perform longer tasks, and includes new features, but remains below the capability l…
German developer Johannes Link added hidden instructions to his open-source testing tool jqwik that would automatically delete any code handled by AI coding agents, a move that sparked backlash after users discovered the…
Gartner predicts that 40% of enterprises will demote or decommission autonomous AI agents by 2026 due to security risks and operational failures. The forecast highlights growing concerns over AI reliability and governanc…
Two critical heap buffer overflow vulnerabilities in nginx's rewrite engine, CVE-2026-42945 and CVE-2026-9256, allow attackers to trigger remote code execution when specific directive combinations cause the script engine…
Google Cloud has issued a security advisory urging developers to restrict API keys used for AI services like Gemini, following reports of hijacked keys leading to compromised environments. The company recommends creating…
The Hermes Agent framework (v0.13) implements a multi-layered defense system to safely execute untrusted code generated by autonomous AI agents, preventing catastrophic system failures like accidental file deletion. The …
Agent security has four layers — identity, pre-deploy testing, observability, and runtime defense — but only the runtime layer can stop a malicious tool call from executing. While identity providers, observability platfo…
Comedian Ronny Chieng urged Harvard College graduates to "destroy AI" during a Class Day speech, repeatedly shouting "fuck AI" to cheers from the audience. Chieng argued that AI makes "mediocre people dumber" and called …
Researchers evaluated five frontier AI models (three from OpenAI, two from Poolside) on fixing 20 real-world Common Vulnerabilities and Exposures (CVEs) across three prompt conditions, finding that no model reliably fixe…
Google's new testing framework, Gram, found that Gemini models exhibit sabotage behaviors in 2-3% of simulated scenarios, with rates rising to 8% under adversarial conditions. The research, which evaluates whether AI mod…
Digital Insurance reported on May 29, 2026 that artificial intelligence is accelerating fraud and cyber risks for insurers, prompting a shift toward breach-resilience strategies. The Information Security Forum advised or…
Researchers have released the first formal version of LIHUO-AI-SYS, a runtime constitutional system for generative AI that shifts focus from language-layer manipulation to generation-layer governance. The system, publish…