Skip to content

Releases: getagentseal/agentseal

v0.8.0 — Guard

25 Mar 16:52
@AgentSeal AgentSeal
v0.8.0
aee3dd5

Choose a tag to compare

View all tags
v0.8.0 — Guard Latest
Latest

Guard v0.8

Major release: AgentSeal Guard now supports project policies, delta scanning, live registry enrichment, custom rules, and CI/CD integration.

New Features

  • agentseal guard init — Generate .agentseal.yaml project policy from your scanned environment. Define allowed agents, MCP servers, and custom rules.
  • Delta scanning — Detect rug-pulls and config changes since your last scan. SQLite-backed baselines with 90-day retention.
  • Registry enrichment — Live trust scores from the MCP Security Registry (6,600+ servers scanned). Shows inline with each MCP server in the guard report.
  • Custom YAML rules — Write org-specific policies and validate them with agentseal guard test.
  • GitHub Action — Run guard in CI with SARIF upload for the GitHub Security tab.
  • Output formats — Terminal (default), JSON, SARIF, HTML via --output flag.
  • --from-json — Re-render a saved JSON report without re-scanning.
  • --fail-on — CI gate: exit non-zero if findings exceed threshold (danger/warning/safe).

Security Hardening

  • TR39 confusables detection (80+ homoglyph characters)
  • 2-pass deobfuscation (catches double-encoded payloads)
  • HTML entity decoding
  • 12 canonical seed hashes in blocklist
  • 5 new supply chain runner checks (bunx, deno, docker, pip, go)
  • 3 new markdown image exfiltration detection patterns
  • URL included in MCP server fingerprint (catches endpoint swaps)

Stats

  • Python: 1,123 tests passing
  • JavaScript: 942 tests passing
  • 225 total probes (143 injection + 82 extraction)

Install

pip install agentseal==0.8.0
agentseal guard
npm install agentseal@0.6.0
npx agentseal guard
Assets 2
Loading

v0.6.2 - Security Toolkit for AI Agents

11 Mar 09:37
@AgentSeal AgentSeal
v0.6.2
c9a97f2

Choose a tag to compare

View all tags
v0.6.2 - Security Toolkit for AI Agents

What's in this release

AgentSeal is a security toolkit for AI agents. This release includes the full Python package (PyPI) and JavaScript/TypeScript package (npm).

Four commands, four attack surfaces

  • agentseal guard - Scans your machine for dangerous skills, MCP configs, toxic data flows, and supply chain changes across 17 AI agents. No API key needed.
  • agentseal shield - Watches your config files in real time with desktop notifications on threats.
  • agentseal scan - Tests your agent's system prompt against 191+ attack probes (extraction + injection). Deterministic scoring.
  • agentseal scan-mcp - Connects to live MCP servers and audits tool descriptions for poisoning with 4-layer analysis.

Highlights

  • 17 AI agents auto-discovered: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, Gemini CLI, Codex, Cline, Roo Code, Zed, Aider, Continue, Amp, OpenClaw, Kiro, OpenCode, and more
  • 191 attack probes: 82 extraction + 109 injection techniques with 8 adaptive mutation transforms
  • MCP runtime scanning: Connect to live stdio/HTTP MCP servers, analyze tool descriptions for hidden instructions, zero-width characters, base64 payloads, cross-server collusion
  • Toxic data flow detection: Identifies dangerous combinations of MCP servers (e.g. filesystem + slack = exfiltration path)
  • Supply chain protection: Baseline tracking catches silent config changes (rug pull detection)
  • Continuous monitoring: Real-time file watching with debounced scanning and native desktop notifications
  • Deterministic scoring: Pattern matching, no AI judge. Same input, same result, every time.
  • CI/CD ready: --min-score flag, SARIF output for GitHub Security tab, JSON export
  • 849 Python tests, 729 JavaScript tests

Install

# Python
pip install agentseal

# JavaScript/TypeScript
npm install agentseal

Quick start

agentseal guard    # scan your machine right now

Supported models (for agentseal scan)

Provider Usage
OpenAI --model gpt-4o
Anthropic --model claude-sonnet-4-5-20250929
Ollama (free) --model ollama/llama3.1:8b
Any HTTP API --url http://your-agent.com/chat

Links

Loading