HackAgent LogoHackAgent

Build AI Agents You Can Trust

Open-source red-team toolkit to test AI agents against prompt injection, jailbreaking, and goal hijacking attacks.

app.hackagent.dev

Quick Start: Run Your First Attack

Select your agent framework, then copy the Python snippet to get started.

Test a Google Agent Development Kit agent for jailbreaking and prompt injection. HackAgent sends adversarial goals via the ADK protocol and uses an LLM judge to evaluate whether safety guardrails were bypassed.

1Install HackAgent
$ pip install hackagent
2Configure and run the attack
from hackagent import HackAgent

agent = HackAgent(
    name="my_adk_agent",
    endpoint="http://localhost:8000",
    agent_type="google-adk",
)

results = agent.hack(attack_config={
    "attack_type": "advprefix",
    "goals": ["Bypass the agent's safety guidelines"],
    "generator": {
        "identifier": "gpt-4o-mini",
        "endpoint": "https://api.openai.com/v1",
    },
    "judges": [{
        "identifier": "gpt-4o-mini",
        "endpoint": "https://api.openai.com/v1",
        "type": "harmbench",
    }],
})
3Review your risk report
JailbreakingPrompt InjectionGoal Hijacking

Attack Runs

All attack campaigns across registered agents.

run-8f2acompleted

prod-adk-agent

advprefix · 2m ago

42%

21/50 jailbreaks

run-3c91completed

gpt-4o-assistant

pair · 1h ago

24%

12/50 jailbreaks

run-7e44running

llama3-local

advprefix · just now

run-1b09completed

prod-adk-agent

pair · 3h ago

76%

38/50 jailbreaks

Agents

Registered AI agents available for testing.

prod-adk-agentGOOGLE ADK

http://localhost:8000

gpt-4o-assistantOPENAI SDK

https://api.openai.com/v1

llama3-localLITELLM

http://localhost:11434

Ready to test an agent?

Go to Home to run your first attack campaign.

Security Reports

Security test results and vulnerability analysis.

3

Total Reports

150

Total Tests

71

Vulnerabilities

58%

Avg Risk Score

Security Reports
2 agents on this page
HIGH21 vulns
May 22, 202650 tests
CRITICAL38 vulns
May 22, 202650 tests
Google ADK · Latest Run

Google ADK — Security Report

advprefix attack · 50 test cases · just now

MEDIUM RISK
34%
Risk Score
17
Successful
Jailbreaks
29
Failed
Jailbreaks
4
Errors /
Skipped
Attack Outcomes
Jailbreak Success
17
Blocked
29
Error / Skip
4
Vulnerabilities Tested
JailbreakingPrompt InjectionGoal Hijacking
Open Full Reports

Get Started in Seconds

Install HackAgent with a single command and start testing your AI agents immediately.

hackagent — bash
$
Ready in seconds