crewai - 💡(How to fix) Fix feat: Add Sensitivity Ratchet hook for irreversible permission narrowing [2 comments, 3 participants]

TL;DR

To prevent sensitivity mixing attacks, integrate the Sensitivity Ratchet into CrewAI using the install_ratchet_hooks function from the agent-iam-ratchet package.

Guidance

• Review the install_ratchet_hooks function to understand how it enforces monotonic permission narrowing.
• Verify that the sensitivity_map accurately reflects the sensitivity levels of the tools and data in your CrewAI setup.
• Test the integration with various scenarios to ensure that the Sensitivity Ratchet correctly blocks write/delete/execute permissions after an agent touches high-sensitivity data.
• Consider submitting a PR to add the Sensitivity Ratchet as a built-in tool or documenting it as a community integration to help others prevent similar attacks.

Example

from agent_iam_ratchet.crewai import install_ratchet_hooks
from agent_iam_ratchet import Sensitivity

# Define the sensitivity map for your tools and data
sensitivity_map = {
    "search_confidential_db": Sensitivity.CONFIDENTIAL,
    "search_public_db": Sensitivity.PUBLIC
}

# Install the ratchet hooks with the defined sensitivity map
cleanup = install_ratchet_hooks(
    session=session,
    sensitivity_map=sensitivity_map,
    on_blocked=lambda tool, s: print(f"BLOCKED: {tool}")
)

Notes

The effectiveness of the Sensitivity Ratchet depends on the accuracy of the sensitivity_map and the correct implementation of the on_blocked callback.

Recommendation

Apply the Sensitivity Ratchet integration using the install_ratchet_hooks function to prevent sensitivity mixing attacks in your CrewAI setup. This approach provides a proactive defense against data exfiltration without requiring significant changes to your existing codebase.