ZIRAN Agent Security Scan
Version updated for https://github.com/taoq-ai/ziran to version v0.12.0.
- This action is used across all versions by 1 repositories.
Action Type
This is a Composite action.
Go to the GitHub Marketplace to find the latest changes.
Action Summary
ZIRAN is a security testing GitHub Action designed to identify vulnerabilities in AI agents, including those using tools, memory, and multi-step reasoning. It automates the discovery of dangerous tool combinations, detects execution-level side effects, and conducts autonomous pentesting campaigns to uncover complex attack paths. Key capabilities include graph-based tool chain analysis, multi-agent coordination, and protocol-specific security testing, making it particularly effective for comprehensive agent security evaluations.
Release notes
What’s Changed
- perf: cache tool classifications to avoid repeated regex matching by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/132
- perf: cache graph state exports and avoid redundant serialization by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/134
- perf: parallelize multi-agent individual scans with asyncio.gather by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/135
- perf: replace O(S*T) find_all_attack_paths with single-source traversal by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/136
- perf: combine refusal detector patterns into single mega-regex by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/137
- perf: deduplicate prompt rendering when applying multiple encodings by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/138
- perf: parallelize protocol auto-detection probes with asyncio.gather by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/139
- perf: bound memory growth of attack results during campaigns by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/141
- fix: prevent import_state() from mutating caller’s dict via .pop() by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/142
- feat: handle HTTP 429 rate limits with Retry-After header support by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/143
- test: add coverage for concurrent attack execution and race conditions by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/156
- test: add coverage for LLM judge timeout and detector pipeline error paths by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/157
- perf: optimize chain analyzer — cache tool nodes, pattern lookups, and centrality by @leoneperdigao in https://github.com/taoq-ai/ziran/pull/133
Full Changelog: https://github.com/taoq-ai/ziran/compare/v0...v0.12.0