agent-iam-ratchet

Sensitivity Ratchet SDK — irreversible permission narrowing for AI agents.

Install

pip install agent-iam-ratchet

With framework integrations:

pip install 'agent-iam-ratchet[langchain]'
pip install 'agent-iam-ratchet[crewai]'
pip install 'agent-iam-ratchet[openai-agents]'

Quick Start

from agent_iam_ratchet import RatchetSession, Sensitivity

session = RatchetSession(scopes=["read:*", "write:*", "delete:*", "execute:*"])

# Agent reads a confidential document
session.access(Sensitivity.CONFIDENTIAL)
print(session.effective_scopes)
# {'read:*', 'write:*'} — delete and execute permanently removed

# Agent reads restricted data
session.access(Sensitivity.RESTRICTED)
print(session.effective_scopes)
# {'read:*'} — only reads remain

# Irreversible — even accessing public data won't restore permissions
session.access(Sensitivity.PUBLIC)
print(session.effective_scopes)
# {'read:*'} — still read-only

Why?

AI agents that access sensitive data can exfiltrate it by writing to lower-sensitivity channels. The sensitivity ratchet prevents this by permanently narrowing permissions after sensitive data access.

License

MIT