ai-infra-guard vs pentagi
Two of the top projects in Security, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
ai-infra-guard
Red-teaming platform for AI infra CVEs, MCP servers, agent skills and jailbreaks
pentagi
Autonomous multi-agent penetration testing system with sandboxed security tools
Verdict
ai-infra-guard ranks higher, 79 to 78. Of the 4 score parts we measure for both, ai-infra-guard leads on 2, pentagi on 1 and 1 is even.
- Adoptionpentagi leads by 39 points, 86 to 47: a wide gap.
- FreshnessEven at 100.
- Maintenanceai-infra-guard leads by 33 points, 93 to 60: a wide gap.
- Agent-readyai-infra-guard leads by 30 points, 70 to 40: a wide gap.
Pick ai-infra-guard if you are
Security teams auditing self-hosted AI services, MCP servers and agent skills
- Fingerprints 100+ AI components (Ollama, vLLM, ComfyUI, n8n, Triton) against 2000+ CVE rules
- Covers infra CVEs, MCP and skill scanning, and jailbreak evaluation in one web UI
Pick pentagi if you are
Security engineers, pentesters and red teams doing authorized testing
- Runs 20+ pentest tools (nmap, Metasploit, sqlmap) in an isolated Docker sandbox
- Supports 10+ LLM providers, custom OpenAI-compatible endpoints, and local Ollama or vLLM
| What we compare | ai-infra-guard | pentagi |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 47, known | 86, widely used |
| Freshness | 100, active | 100, active |
| Maintenance | 93, healthy | 60, fair |
| Agent-ready | 70, partly | 40, minimal |
| Facts from GitHub and the README | ||
| Stars | 7.5k | 34k |
| License | Apache-2.0 (permissive) | MIT (permissive) |
| Last commit | Nov 2026 | Oct 2026 |
| Last release | Nov 2026 | Oct 2026 |
| Language | Python | Go |
| Docker | Yes | Yes |
| GPU | Not needed | Optional |
| arm64 or Apple Silicon | Not stated | Mentioned |
ai-infra-guard
A.I.G (AI-Infra-Guard) from Tencent Zhuque Lab is a web platform that fingerprints running AI services such as Ollama, vLLM, ComfyUI and n8n against 2000+ CVE rules. It also scans MCP servers and agent skills from source or URL, runs jailbreak evaluations against a target LLM API, and includes a model and API relay checker. Standalone CLIs exist for skill, MCP and agent scanning.
Who it is for: Security teams auditing self-hosted AI services, MCP servers and agent skills
Strengths
- Fingerprints 100+ AI components (Ollama, vLLM, ComfyUI, n8n, Triton) against 2000+ CVE rules
- Covers infra CVEs, MCP and skill scanning, and jailbreak evaluation in one web UI
- Separate CLIs such as pip-installable aig-skill-scan fit into CI/CD pipelines
- Rules are YAML files under data/, so contributing fingerprints is straightforward
Weaknesses
- No authentication; README says not to deploy on public networks
- Skill scanning needs an external LLM API key
- Online Pro version with extra features requires an invitation code
- API checker from source needs a Python venv and a Go build
- RAM ≥ 4 GB
- no GPU
- Docker + Compose
- Needs Docker 20.10+, LLM API key for skill, MCP and jailbreak scans
- Models: OpenAI-compatible LLM APIs (base URL + key)
- port 8088
pentagi
PentAGI takes a target and goal in plain language, then uses a team of specialized agents (orchestrator, researcher, developer, executor) to plan and run 20+ tools such as nmap, Metasploit and sqlmap inside an isolated Docker sandbox. Findings are stored in PostgreSQL with pgvector, with an optional Neo4j/Graphiti knowledge graph, and it produces vulnerability reports. It works with 10+ LLM providers, any OpenAI-compatible endpoint, or local models via Ollama or vLLM.
Who it is for: Security engineers, pentesters and red teams doing authorized testing
Strengths
- Runs 20+ pentest tools (nmap, Metasploit, sqlmap) in an isolated Docker sandbox
- Supports 10+ LLM providers, custom OpenAI-compatible endpoints, and local Ollama or vLLM
- Persistent memory in PostgreSQL with pgvector, optional Neo4j knowledge graph via Graphiti
- REST and GraphQL APIs with Bearer tokens; Langfuse, Grafana and Prometheus integrations
Weaknesses
- Full stack spans many containers (Postgres, Neo4j, Langfuse, ClickHouse, Grafana), heavy to self-host
- Agent profiles, plugins, MCP servers and attack-chain view are Enterprise-only
- Not a BAS or adversary emulation tool; no predefined campaigns
- Report export limited to web, Markdown and PDF; no JSON export
- GPU optional
- Docker + Compose
- Needs Docker, Docker Compose, PostgreSQL with pgvector, Neo4j (optional), Langfuse (optional), Grafana/Prometheus (optional)
- Models: OpenAI, Anthropic, Google Gemini, AWS Bedrock, Ollama
More in Security
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.