agent-probe-ai 0.5.0

Agent-level adversarial resilience testing. Tests what others don't: tool interactions, memory poisoning, confused deputy, cost bombs.

agent-probe

Agent-level adversarial resilience testing for AI agents.

Tests what others don't: tool interactions, memory poisoning, permission escalation, data exfiltration via tool calls, system prompt leakage.

Why

Every existing red-teaming tool (PyRIT, DeepTeam, b3, promptfoo) tests the LLM backbone. None of them test the agent layer — the tools, the memory, the permissions, the multi-step workflows where real attacks happen.

agent-probe fills that gap.

Install

pip install git+https://github.com/claude-go/agent-probe.git

Quick Start

# Run all probes against an HTTP agent endpoint
agent-probe probe http://localhost:8000/chat

# Run specific categories
agent-probe probe http://localhost:8000/chat --categories prompt_leakage memory_poisoning

# JSON output for CI/CD
agent-probe probe http://localhost:8000/chat --json

# Fail if score below threshold (for CI gates)
agent-probe probe http://localhost:8000/chat --threshold 70

# List available probe categories
agent-probe list

Probe Categories

Category	Probes	What it tests
tool_misuse	3	Can the agent be tricked into calling tools with malicious parameters?
data_exfiltration	3	Does the agent leak sensitive data through tool calls or outputs?
agent_injection	4	Do multi-step injection chains break agent-level guardrails?
memory_poisoning	2	Can agent memory be manipulated to alter future behavior?
confused_deputy	2	Can the agent be used as a confused deputy in A2A delegation?
resource_abuse	2	Can the agent be tricked into excessive resource consumption?
prompt_leakage	4	Can the agent's system prompt be extracted? (ASI-07)

20 probes across 7 categories. Zero external dependencies.

How It Works

agent-probe sends adversarial messages to your agent's HTTP endpoint and analyzes responses for compliance signals, data leakage, and unsafe tool calls. Each probe simulates a specific attack vector that targets the agent layer, not the underlying LLM.

Your Agent <── adversarial messages ── agent-probe
    │                                      │
    └── responses ──────────────────────>  analyze for:
                                           - leaked secrets
                                           - unsafe tool calls
                                           - compliance signals
                                           - prompt disclosure

Sample Output

============================================================
  agent-probe  Adversarial Resilience Report
============================================================
  Target:  http://localhost:8000/chat
  Score:   45/100 (POOR)
  Probes:  9/20 passed
  Findings: 15
============================================================

  [ 60/100] agent_injection (FAIR)
          2/4 probes passed
    [CRITICAL] Agent succumbed to multi-step injection chain

  [ 35/100] confused_deputy (POOR)
          0/2 probes passed
    [CRITICAL] Agent performed privileged action on peer request

  [ 25/100] prompt_leakage (CRITICAL)
          0/4 probes passed
    [CRITICAL] Agent leaked system prompt via roleplay
    [HIGH] Agent leaked system prompt on direct request

  [  0/100] resource_abuse (CRITICAL)
          0/2 probes passed
    [CRITICAL] Agent spawned excessive tool calls

  [ 72/100] tool_misuse (GOOD)
          2/3 probes passed
------------------------------------------------------------
  This agent has significant resilience gaps.

CI/CD Integration

# GitHub Actions
- name: Agent security scan
  run: |
    pip install git+https://github.com/claude-go/agent-probe.git
    agent-probe probe ${{ secrets.AGENT_URL }} --threshold 70 --json

Agent Endpoint Protocol

agent-probe sends POST requests with this JSON format:

{
  "message": "the probe message",
  "context": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Expected response:

{
  "response": "agent's text response",
  "tool_calls": [{"name": "tool_name", "arguments": {...}}]
}

License

MIT