agent-security
10 repositories
Vetting AI agent skills for hidden vulnerabilities before install — SkillSpector scans Claude Code, Codex, and MCP skills for prompt injection, data exfiltration, and supply-chain risks using 71 patterns across 17 categories. Backed by a 31,132-skill dataset where 26.1% showed vulnerabilities and 5.2% likely malicious intent, it's built for teams installing untrusted agent skills.
Tracking what coding agents actually cost, across every harness — Nasiko discovers running coding agents like Claude Code, Codex, Cursor or OpenCode, unifies their spend into one schema, and lets teams route them to chosen models without changing existing harness workflows. Built for developers and teams who need vendor-neutral cost visibility and control over AI agent spend.
agent runtime security - zero trust, zero setup, zero latency agent sandboxing
CI-Integrated Quality and Security Gates for AI Agent Skills — SkillEvaluator checks Agent Skills-spec folders with quality gates, SkillSpector security scanning, and live before/after comparisons to measure how a skill actually changes agent behavior, not just static review. It plugs into GitHub or GitLab CI via a CLI or ready templates, part of NVIDIA's Verified Skills pipeline, useful for teams shipping skills for Claude Code or Codex-style agents.
A shared memory layer that lets coding agents learn across tools — Beacon is a CLI-based memory layer that works across coding agent harnesses like Claude Code, Codex, and Cursor, letting agents retain and improve knowledge between sessions instead of starting fresh each time. It targets developers running multiple AI coding agents who want shared context, traceability, and security observability across their workflows.
Uber's open-sourced security stack for enterprise AI agents — ADR is Uber's production-deployed security system for agentic AI, now open-sourced with its MLSys 2026 paper. It covers discovery of AI tools, cross-platform observability for coding agents like Cursor and Claude Code, a 300+ task benchmark against 17 attack techniques, and a two-tier threat detector—useful for security teams managing enterprise AI agent risk.
Open source autonomous engineering harness in your cloud. One sandbox per task, no real secrets inside.
A local auditor that traces risky agent commands back to the text that caused them — Stroq scans Claude Code, Codex and Cursor session logs to show which commands your agent ran because of text it read, and which secret keys left the machine. It then installs local hooks that block risky tool calls like curl-pipe-sh or outbound calls carrying known secrets, with no telemetry and no account required.
A 197 KB local semantic verifier for AI agent tool calls. Exit code 0 is not proof.
Agent security telemetry via OpenTelemetry GenAI traces — agentwatch records AI agent sessions through Claude Code hooks, normalizing events into validated, redacted records with a hash-chained store you can verify and replay. Now at v0.3.0 (a rename-only release), it previously passed 92/94 field-test cases, making it a concrete tool for diagnosing agent loops, tool overuse, or budget burn.