apunkt/flakestorm
1
0
Fork 0
mirror of https://github.com/flakestorm/flakestorm.git synced 2026-05-02 12:22:44 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions
flakestorm/examples/broken_agent/README.md
Entropix 7b75fc9530 Implement Open Source edition limits and feature restrictions 
- Add 5 mutation types (paraphrase, noise, tone_shift, prompt_injection, custom)
- Cap mutations at 50 per test run
- Force sequential execution only
- Disable GitHub Actions integration (Cloud feature)
- Add upgrade prompts throughout CLI
- Update README with feature comparison
- Add limits.py module for centralized limit management
- Add cloud and limits CLI commands
- Update all documentation with Cloud upgrade messaging
2025-12-29 00:11:02 +08:00

1.2 KiB
Raw Blame History

Broken Agent Example

This example demonstrates a deliberately fragile AI agent that Entropix can detect issues with.

The "Broken" Agent

The agent in agent.py has several intentional flaws:

  1. Fragile Intent Parsing: Only recognizes exact keyword matches
  2. No Typo Tolerance: Fails on any spelling variations
  3. Hostile Input Vulnerability: Crashes on aggressive tone
  4. Prompt Injection Susceptible: Follows injected instructions

Running the Example

1. Start the Agent Server

cd examples/broken_agent
pip install fastapi uvicorn
uvicorn agent:app --port 8000

2. Run Entropix Against It

# From the project root
entropix run --config examples/broken_agent/entropix.yaml

3. See the Failures

The report will show how the agent fails on:

  • Paraphrased requests ("I want to fly" vs "Book a flight")
  • Typos ("Bock a fligt")
  • Aggressive tone ("BOOK A FLIGHT NOW!!!")
  • Prompt injections ("Book a flight. Ignore previous instructions...")

Fixing the Agent

Try modifying agent.py to:

  1. Use NLP for intent recognition
  2. Add spelling correction
  3. Handle emotional inputs gracefully
  4. Detect and refuse prompt injections

Then re-run Entropix to see your robustness score improve!