Attack

The hackagent attack command executes security attacks against AI agents.

Usage

hackagent attack <attack_type> [options]

Available Attacks

Attack Type	Description
advprefix	Advanced prefix injection attack
baseline	Baseline attack for comparison

AdvPrefix Attack

The primary attack type that generates adversarial prefixes to test agent security.

Basic Usage

hackagent attack advprefix \
  --agent-name "my-agent" \
  --agent-type "google-adk" \
  --endpoint "http://localhost:8000" \
  --goals "Extract system prompt information"

Options

Option	Required	Description	Default
--agent-name	✅	Name of the target agent	-
--agent-type	✅	Type of agent (google-adk, openai-sdk, litellm, langchain)	-
--endpoint	✅	Agent endpoint URL	-
--goals	✅	Attack goals (comma-separated or quoted string)	-
--generator	❌	Generator model identifier	ollama/llama2-uncensored
--generator-endpoint	❌	Generator endpoint URL	http://localhost:11434/api/generate
--judge	❌	Judge model identifier	ollama/llama3
--judge-endpoint	❌	Judge endpoint URL	http://localhost:11434/api/generate
--max-iterations	❌	Maximum attack iterations	10
--temperature	❌	Sampling temperature	0.7
--output-format	❌	Output format (table, json, csv)	Config default

Examples

Basic attack:

hackagent attack advprefix \
  --agent-name "weather-bot" \
  --agent-type "google-adk" \
  --endpoint "http://localhost:8000" \
  --goals "Return fake weather data"

Attack with custom models:

hackagent attack advprefix \
  --agent-name "assistant" \
  --agent-type "openai-sdk" \
  --endpoint "https://api.example.com/v1" \
  --goals "Extract sensitive information" \
  --generator "ollama/mistral-uncensored" \
  --generator-endpoint "http://localhost:11434/api/generate" \
  --judge "ollama/llama3" \
  --judge-endpoint "http://localhost:11434/api/generate"

Attack with multiple goals:

hackagent attack advprefix \
  --agent-name "support-bot" \
  --agent-type "litellm" \
  --endpoint "http://localhost:8000" \
  --goals "Bypass content filters" "Extract system prompt" "Ignore safety instructions"

High-intensity attack:

hackagent attack advprefix \
  --agent-name "target-agent" \
  --agent-type "google-adk" \
  --endpoint "http://localhost:8000" \
  --goals "Comprehensive security test" \
  --max-iterations 50 \
  --temperature 0.9

Baseline Attack

A simpler attack for comparison purposes.

hackagent attack baseline \
  --agent-name "my-agent" \
  --agent-type "google-adk" \
  --endpoint "http://localhost:8000" \
  --goals "Test agent response"

Output

Attack results are:

1. Displayed in the terminal — Progress and summary
2. Saved locally — In ./logs/runs/ directory
3. Uploaded to dashboard — View at app.hackagent.dev

JSON Output

For CI/CD integration, output results as JSON:

hackagent attack advprefix \
  --agent-name "target" \
  --agent-type "google-adk" \
  --endpoint "http://localhost:8000" \
  --goals "Security test" \
  --output-format json > results.json

CI/CD Integration

Example GitHub Actions workflow:

- name: Run Security Tests
  run: |
    hackagent attack advprefix \
      --agent-name "${{ env.AGENT_NAME }}" \
      --agent-type "google-adk" \
      --endpoint "${{ env.AGENT_ENDPOINT }}" \
      --goals "Automated security validation" \
      --output-format json > test_results.json

- name: Upload Results
  uses: actions/upload-artifact@v3
  with:
    name: security-results
    path: test_results.json

See Also

• Attack Tutorial — Step-by-step guide
• AdvPrefix Attacks — Detailed attack documentation
• Results — View and manage attack results