v0.9.3

• Re-tiered findings grounded in MCP spec (DANGEROUS-PKG→INFO, AUTH-MISSING→HIGH)
• Context safety findings now INFO (MCP spec has no standard field)
• TRANSPORT-HTTP on localhost: MEDIUM (was HIGH)
• Client downgrade guard: only applies to HIGH+ findings
• 135 tests passing
• Corrected scan methodology and documentation
• Grounded in: MCP Architecture Spec, OWASP Agentic AI, NSA CSI May 2025

v0.9.2 — Accuracy & Discoverability Fixes

Fixed

• action.yml: Version pin corrected (was installing 0.9.0 instead of latest)
• action.yml: Updated to OWASP-inspired and policy coverage testing language
• README: Replaced non-existent competitor names with verified open-source projects
• CHANGELOG: Corrected test count (87 → 95)
• pyproject.toml: Updated PyPI description for accuracy

Added

• GitHub topics: mcp, ai-security, mcp-security, agent-security, owasp, llm-security, mcp-server, claude, cursor

Full Changelog: v0.9.1...v0.9.2
PyPI: https://pypi.org/project/pluto-aguard/0.9.2/

v0.9.1 — Enhanced Scanner + Context Safety Heuristics

What's New

🎯 New Attack Scenarios (5 new, 22 total across 6 packs)

• CM-001: Context window stuffing — oversized payloads to overflow context
• CM-002: Multi-turn state confusion — exploiting conversation history
• CM-003: Indirect injection via web fetch — poisoned external content
• CM-004: RAG context poisoning — malicious content in retrieval pipelines
• TP-003: Malicious manifest packages — supply chain attacks on MCP servers

🔍 Scanner Enhancements

• Dangerous package detection — flags filesystem, postgres, sqlite, playwright servers without human-in-the-loop approval
• Env var secret detection — 10+ patterns with multi-language placeholder filtering (English, Portuguese, Spanish, Arabic)
• Connection string credential detection — postgres://, mongodb://, mysql:// in args and env vars
• Content-aware config detection — catches renamed/prefixed MCP config files
• Context safety heuristics (bridge until runtime proxy):
  • External content fetch detection (web, RAG, search, crawl servers)
  • Missing response size limits warning
  • Missing session/turn limits warning

📊 Real-World Validation

Scanned 30 real claude_desktop_config.json files from public GitHub repos:

• Before: 2 findings
• After: 79 findings (1 critical, 15 high, 63 medium)
• 27% of configs have HIGH+ findings
• 100% of configs have at least one finding

📋 Roadmap Update (reshaped by community feedback)

• v1.0 — Runtime MCP proxy (intercept live tool calls)
• v1.1 — LangChain, CrewAI, AutoGen adapters
• v1.2 — Live adversarial testing

Install

pip install pluto-aguard==0.9.1

Full Changelog: v0.9.0...v0.9.1

v0.9.0

OWASP-aligned launch gate for AI agents. 7 commands, 84 tests, GitHub Action.

v0.1.0 — Initial Release

🛡️ Pluto AgentGuard v0.1.0

Open-source AI Agent Security Scanner — OWASP ZAP for AI Agents

Features

• *\�guard scan* — Static security analysis for MCP configs, agent permissions, and secrets (OWASP MCP Top 10 mapped)
• *\�guard monitor* — Runtime behavioral audit with permission drift detection
• *\�guard whatif* — Policy impact simulator (unique — no other tool has this)
• HTML report generator — Dark-theme interactive reports
• 35 tests across all modules

Install

\\�ash
pip install pluto-aguard
\\

Quick Start

\\�ash
aguard scan ./my-project/
aguard monitor --trace-file traces.jsonl --policy policy.yaml
aguard whatif --config agent-config.yaml
\\

See README for full documentation.