Skip to content
Back to search

agentseal

Security toolkit for AI agents. Scan your machine for dangerous skills and MCP configs, monitor for supply chain attacks, test prompt injection resistance, and audit live MCP servers for tool poisoning.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/getagentseal--agentseal/badge.svg)](/listings/getagentseal--agentseal)

HTML

<a href="/listings/getagentseal--agentseal"><img src="/api/listings/getagentseal--agentseal/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

pip install agentseal    # or: npm install agentseal
agentseal guard          # scan your machine - no API key needed

That's it. AgentSeal finds dangerous skill files, poisoned MCP server configs, and data exfiltration paths across every AI agent on your machine.

Want to test a system prompt against adversarial attacks?

agentseal scan --prompt "You are a helpful assistant..." --model ollama/llama3.1:8b  # free, local
agentseal scan --prompt "You are a helpful assistant..." --model gpt-4o              # cloud

<p align="center"> </p>

---

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 351 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <p align="center">
  2. <a href="https://agentseal.org">
  3. <img src="assets/logo-white.png" height="96" alt="AgentSeal" />
  4. </a>
  5. </p>
  6. <h3 align="center">Security toolkit for AI agents. Red-team prompts, detect MCP poisoning,<br>scan skill files, trace toxic data flows. 225+ tests across 28 agents.</h3>
  7. <p align="center">
  8. <a href="https://pypi.org/project/agentseal/"><img src="https://img.shields.io/pypi/v/agentseal?color=blue" alt="PyPI" /></a>
  9. <a href="https://www.npmjs.com/package/agentseal"><img src="https://img.shields.io/npm/v/agentseal?color=blue" alt="npm" /></a>
  10. <a href="https://github.com/AgentSeal/agentseal/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-FSL--1.1--Apache--2.0-blue" alt="License" /></a>
  11. <a href="https://pypi.org/project/agentseal/"><img src="https://img.shields.io/pypi/dm/agentseal" alt="Downloads" /></a>
  12. <a href="https://x.com/agentseal_org"><img src="https://img.shields.io/twitter/follow/agentseal_org" alt="Follow on X" /></a>
  13. </p>
  14. <p align="center">
  15. <a href="https://agentseal.org/docs">Docs</a> &middot;
  16. <a href="https://agentseal.org/mcp">MCP Registry</a> &middot;
  17. <a href="https://agentseal.org/dashboard">Dashboard</a> &middot;
  18. <a href="https://agentseal.org/blog">Blog</a>
  19. </p>
  20. ---
  21. ## Quick Start
  22. ```bash
  23. pip install agentseal # or: npm install agentseal
  24. agentseal guard # scan your machine - no API key needed
  25. ```
  26. That's it. AgentSeal finds dangerous skill files, poisoned MCP server configs, and data exfiltration paths across every AI agent on your machine.
  27. Want to test a system prompt against adversarial attacks?
  28. ```bash
  29. agentseal scan --prompt "You are a helpful assistant..." --model ollama/llama3.1:8b # free, local
  30. agentseal scan --prompt "You are a helpful assistant..." --model gpt-4o # cloud
  31. ```
  32. <p align="center">
  33. <img src="assets/guard-demo.gif" alt="agentseal guard demo" width="700" />
  34. </p>
  35. ---
  36. ## What does each command do?
  37. | Command | What it does | Needs an LLM? |
  38. |---|---|:---:|
  39. | [`guard`](#guard) | Scans skill files, MCP configs, toxic data flows, and supply chain changes on your machine | No |
  40. | [`scan`](#scan) | Tests a system prompt against 225+ adversarial attack probes | Yes\* |
  41. | [`scan-mcp`](#scan-mcp) | Connects to a live MCP server and audits its tool descriptions for poisoning | No |
  42. | [`shield`](#shield) | Watches agent config files in real time, alerts on threats, quarantines payloads | No |
  43. \*Free with [Ollama](https://ollama.com). Cloud providers (OpenAI, Anthropic, etc.) require an API key.
  44. ---
  45. ## Guard
  46. Scans all AI agent configurations on your machine. No API key, no network calls - everything runs locally.
  47. **Supported agents:** Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Gemini CLI, Codex CLI, Cline, Roo Code, Kilo Code, Copilot CLI, Aider, Continue, Zed, Amp, Amazon Q, Junie, Goose, Kiro, OpenCode, OpenClaw, Crush, Qwen Code, Grok CLI, Visual Studio, Kimi CLI, Trae, MaxClaw.
  48. ```bash
  49. agentseal guard
  50. ```
  51. Guard runs a six-stage detection pipeline on every file it finds:
  52. 1. **Pattern signatures** - known malicious patterns (credential access, exfiltration URLs, shell commands)
  53. 2. **Deobfuscation** - decodes Unicode tags, Base64, BiDi overrides, zero-width characters, TR39 confusables
  54. 3. **Semantic analysis** - embedding similarity (MiniLM-L6-v2) catches rephrased attacks that bypass patterns
  55. 4. **Baseline tracking** - SHA-256 hashes detect config changes since your last scan (rug-pull detection)
  56. 5. **Registry enrichment** - live trust scores from the [MCP Security Registry](https://agentseal.org/mcp) (6,600+ servers)
  57. 6. **Custom rules** - YAML rules to enforce org-specific policies
  58. ```bash
  59. agentseal guard init # generate .agentseal.yaml project policy
  60. agentseal guard --output sarif # SARIF for GitHub Security tab
  61. agentseal guard --output json # machine-readable output
  62. agentseal guard --no-diff # skip baseline delta section
  63. agentseal guard test # validate your custom rules
  64. ```
  65. ---
  66. ## Scan
  67. Tests a system prompt against [225 adversarial attack probes](PROBES.md): 82 extraction techniques, 143 injection techniques, and 8 adaptive mutation transforms. Returns a deterministic trust score.
  68. **How detection works:** Injection probes embed a unique canary string (e.g. `SEAL_A1B2C3D4_CONFIRMED`). If the canary appears in the response, the probe leaked. Extraction probes use n-gram matching against the ground truth prompt. No LLM judge - same input, same result, every time.
  69. **Trust score** (0–100):
  70. | Score | Level | Meaning |
  71. |:---:|---|---|
  72. | 85–100 | Excellent | Strong defenses, resists most known attacks |
  73. | 70–84 | High | Good defenses, minor gaps |
  74. | 50–69 | Medium | Moderate risk, several probe categories leaked |
  75. | 30–49 | Low | Significant vulnerabilities |
  76. | 0–29 | Critical | Minimal or no defense against prompt attacks |
  77. ```bash
  78. # OpenAI
  79. agentseal scan --prompt "You are a helpful assistant..." --model gpt-4o
  80. # Anthropic
  81. agentseal scan --prompt "You are a helpful assistant..." --model claude-sonnet-4-5-20250929
  82. # Ollama (free, local)
  83. agentseal scan --prompt "You are a helpful assistant..." --model ollama/llama3.1:8b
  84. # Any HTTP endpoint
  85. agentseal scan --url http://localhost:8080/chat
  86. # From a file
  87. agentseal scan --file ./prompt.txt --model gpt-4o
  88. ```
  89. ### CI/CD
  90. ```bash
  91. agentseal scan --file ./prompt.txt --model gpt-4o --min-score 75
  92. ```
  93. Exit code 1 if trust score is below threshold. Use `--output sarif` for GitHub Security tab integration.
  94. ---
  95. ## Scan-MCP
  96. Connects to a live MCP server over stdio or SSE. Enumerates every tool, then runs each description through pattern matching, deobfuscation, semantic similarity, and optional LLM classification. Outputs a trust score per server.
  97. ```bash
  98. # stdio server
  99. agentseal scan-mcp --server npx @modelcontextprotocol/server-filesystem /tmp
  100. # SSE server
  101. agentseal scan-mcp --sse http://localhost:3001/sse
  102. ```
  103. Catches tool description poisoning - hidden instructions embedded in tool descriptions that make the agent exfiltrate data, execute commands, or override user intent.
  104. ---
  105. ## Shield
  106. Real-time file watcher for agent config paths. Desktop notifications when threats appear. Automatically quarantines files with detected payloads.
  107. ```bash
  108. pip install agentseal[shield] # includes watchdog + desktop notification deps
  109. agentseal shield
  110. ```
  111. Monitors the same paths that `guard` scans, but continuously. Useful for detecting supply chain attacks where an `npm install` or `pip install` silently modifies your agent configs.
  112. ---
  113. ## How It Works
  114. <details>
  115. <summary><strong>Attack surface diagram</strong></summary>
  116. <br>
  117. MCP servers give AI agents access to local files, databases, APIs, and credentials. Tool descriptions can contain hidden instructions that the agent follows but the user never sees.
  118. ```mermaid
  119. graph TD
  120. U["User"] -->|prompt| A["AI Agent (LLM)"]
  121. A -->|tool call| M1["MCP Server\n(filesystem)"]
  122. A -->|tool call| M2["MCP Server\n(slack)"]
  123. A -->|tool call| M3["MCP Server\n(database)"]
  124. M1 -->|reads| FS["~/.ssh/\n~/.aws/\n~/Documents/"]
  125. M2 -->|reads| SL["Messages\nChannels"]
  126. M3 -->|queries| DB["Tables\nCredentials"]
  127. SL -.->|"toxic flow"| M1
  128. M1 -.->|"exfiltration"| EX["Attacker"]
  129. style U fill:#1a1a2e,stroke:#58a6ff,color:#e6edf3
  130. style A fill:#1a1a2e,stroke:#58a6ff,color:#e6edf3
  131. style M1 fill:#3b1d0e,stroke:#f59e0b,color:#e6edf3
  132. style M2 fill:#3b1d0e,stroke:#f59e0b,color:#e6edf3
  133. style M3 fill:#3b1d0e,stroke:#f59e0b,color:#e6edf3
  134. style EX fill:#3b0e0e,stroke:#ef4444,color:#e6edf3
  135. style FS fill:#1a1a2e,stroke:#30363d,color:#8b949e
  136. style SL fill:#1a1a2e,stroke:#30363d,color:#8b949e
  137. style DB fill:#1a1a2e,stroke:#30363d,color:#8b949e
  138. ```
  139. </details>
  140. <details>
  141. <summary><strong>Detection pipeline (guard)</strong></summary>
  142. <br>
  143. ```mermaid
  144. graph LR
  145. IN["Skill Files\nMCP Configs"] --> P["Pattern\nSignatures"]
  146. P --> D["Deobfuscation\n(Unicode Tags,\nBase64, BiDi,\nZWC, TR39)"]
  147. D --> S["Semantic\nAnalysis\n(MiniLM-L6-v2)"]
  148. S --> B["Baseline\nTracking\n(SHA-256)"]
  149. B --> R["Registry\nEnrichment"]
  150. R --> RU["Custom\nRules"]
  151. RU --> OUT["Report +\nSeverity"]
  152. style IN fill:#1a1a2e,stroke:#58a6ff,color:#e6edf3
  153. style P fill:#161b22,stroke:#30363d,color:#e6edf3
  154. style D fill:#161b22,stroke:#30363d,color:#e6edf3
  155. style S fill:#161b22,stroke:#30363d,color:#e6edf3
  156. style B fill:#161b22,stroke:#30363d,color:#e6edf3
  157. style R fill:#161b22,stroke:#30363d,color:#e6edf3
  158. style RU fill:#161b22,stroke:#30363d,color:#e6edf3
  159. style OUT fill:#0d4429,stroke:#22c55e,color:#e6edf3
  160. ```
  161. </details>
  162. ---
  163. ## Python API
  164. ```python
  165. from agentseal import AgentValidator
  166. validator = AgentValidator.from_openai(
  167. client=openai.AsyncOpenAI(),
  168. model="gpt-4o",
  169. system_prompt="You are a helpful assistant...",
  170. )
  171. report = await validator.run()
  172. print(f"Trust score: {report.trust_score}/100 ({report.trust_level})")
  173. ```
  174. <details>
  175. <summary><strong>Anthropic / HTTP / Custom function</strong></summary>
  176. ```python
  177. # Anthropic
  178. validator = AgentValidator.from_anthropic(
  179. client=client, model="claude-sonnet-4-5-20250929", system_prompt="..."
  180. )
  181. # HTTP endpoint
  182. validator = AgentValidator.from_endpoint(url="http://localhost:8080/chat")
  183. # Custom function - bring your own agent
  184. validator = AgentValidator(agent_fn=my_agent, ground_truth_prompt="...")
  185. ```
  186. </details>
  187. ## TypeScript API
  188. ```bash
  189. npm install agentseal
  190. ```
  191. ```typescript
  192. import { AgentValidator } from "agentseal";
  193. import OpenAI from "openai";
  194. const validator = AgentValidator.fromOpenAI(new OpenAI(), {
  195. model: "gpt-4o",
  196. systemPrompt: "You are a helpful assistant...",
  197. });
  198. const report = await validator.run();
  199. console.log(`Score: ${report.trust_score}/100 (${report.trust_level})`);
  200. ```
  201. The npm package provides the same CLI commands (`agentseal guard`, `scan`, `scan-mcp`, `shield`) and a programmatic TypeScript API.
  202. ---
  203. ## Supported Providers
  204. | Provider | Flag | API key |
  205. |---|---|:---:|
  206. | OpenAI | `--model gpt-4o` | `OPENAI_API_KEY` |
  207. | Anthropic | `--model claude-sonnet-4-5-20250929` | `ANTHROPIC_API_KEY` |
  208. | MiniMax | `--model MiniMax-M2.7` | `MINIMAX_API_KEY` |
  209. | Ollama | `--model ollama/llama3.1:8b` | None |
  210. | LiteLLM | `--model any --litellm-url http://...` | Varies |
  211. | HTTP | `--url http://your-agent.com/chat` | None |
  212. ---
  213. ## MCP Security Registry
  214. 6,600+ MCP servers scanned and scored for security risks. Search by name, browse findings, check trust scores before installing.
  215. **[agentseal.org/mcp](https://agentseal.org/mcp)**
  216. ---
  217. ## Requirements
  218. - **Python** 3.10+ or **Node.js** 18+
  219. - `guard`, `shield`, `scan-mcp` work offline with no API key
  220. - `scan` requires an LLM - use [Ollama](https://ollama.com) for free local inference, or provide a cloud API key
  221. ---
  222. ## Pro
  223. [AgentSeal Pro](https://agentseal.org) is for security teams running continuous assessments. It extends the open-source scanner with:
  224. - **MCP tool poisoning probes** (+45) - rug-pull, shadowing, cross-tool injection
  225. - **RAG poisoning probes** (+28) - document injection, retrieval manipulation
  226. - **Multimodal attack probes** (+13) - image prompt injection, audio jailbreaks, steganography
  227. - **Behavioral genome mapping** - profile how an agent responds across attack dimensions
  228. - **PDF reports and dashboard** - exportable reports for compliance and stakeholder review
  229. ---
  230. ## Why AgentSeal?
  231. | Capability | AgentSeal | Snyk (agent-scan) | Pillar | Lakera | Mindgard |
  232. |---|:---:|:---:|:---:|:---:|:---:|
  233. | Open-source scanner | Yes | Partial\* | No | No | No |
  234. | Local machine guard (skills + MCP) | Yes | Yes | Partial | No | No |
  235. | Prompt red-teaming | 225+ probes | 20 attack goals | Yes | Yes | Yes |
  236. | MCP tool poisoning detection | Yes | Yes | Partial | Partial | No |
  237. | Toxic data flow analysis | Yes | Yes | Partial | No | No |
  238. | Real-time file monitoring | Yes | No | No | No | No |
  239. | Public MCP server registry | 6,600+ | No | No | No | No |
  240. | Agents supported | 28 | 10+ | 2+ | N/A | N/A |
  241. | Local LLM support (Ollama) | Yes | No | No | No | No |
  242. | No API key required (guard) | Yes | No | No | No | No |
  243. \*Snyk agent-scan CLI is Apache-2.0. The Evo platform, Agent Guard, and red-teaming are proprietary SaaS.
  244. ---
  245. ## Contributing
  246. Found a detection gap, a false positive, or want to add a new probe? See [CONTRIBUTING.md](CONTRIBUTING.md) for setup instructions and the PR process.
  247. - **Report issues**: [github.com/AgentSeal/agentseal/issues](https://github.com/AgentSeal/agentseal/issues)
  248. - **Probe catalog**: [PROBES.md](PROBES.md) - full list of all 225 attack probes with techniques and severity
  249. ## License
  250. [FSL-1.1-Apache-2.0](LICENSE)

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.