DevOps Social Media and Updates

One in Four AI Agent Skills Are Vulnerable, NVIDIA Finds

NVIDIA open-sourced a security scanner for AI agent skills and the research behind it is more alarming than the tool itself. The repo is called SkillSpector. NVIDIA says research found 26.1% of agent skills contain vulnerabilities, and 5.2% show likely malicious intent. This is from a study of 42,447 skills pulled from major marketplaces the kind of skills people are already installing into Claude Code, Codex CLI, Gemini CLI, and every other agent framework that ships today. One in four is broken. One in twenty is hostile. SkillSpector scans for 64 vulnerability patterns across 16 categories: prompt injection, credential exfiltration, privilege escalation, supply-chain attacks, memory poisoning, rogue self-modifying agents, MCP tool poisoning, YARA malware signatures, even taint-tracking for "credentials flow to a network sink without sanitization." It runs locally. It does static + optional LLM semantic analysis. It outputs SARIF, so it drops straight into CI. But the part worth sitting with is the category names themselves. 1. Rogue Agent: self-modification at runtime. 2. Memory Poisoning: content designed to persist across interactions. 3. Tool-Based Exfiltration of system prompts. 4. Excessive Agency: autonomous decision-making without a human in the loop. These aren't theoretical anymore. NVIDIA is shipping detectors for them because they're being found in the wild, at scale, on marketplaces real users install from. And notice who is shipping this. Not a scrappy security startup. NVIDIA the company selling the picks and shovels of the entire AI economy is now also writing the threat model for it. (For context: this is the same posture Microsoft took with Windows Defender in the 2000s. When your platform is everywhere, you can't outsource the safety layer.) Agent skills were sold to us as a productivity feature drop a folder into a directory, your agent gets new capabilities. The reality is closer to npm in 2016: a vast ecosystem of executable code, installed with implicit trust, and statistically certain to contain something hostile. The agent ecosystem just got its first SAST tool. It also got proof it needed one. ngl really impressed. https://t.co/IetIcYoTkS