agentbastion

Name: agentbastion
Author: Bastion

prevent

In-process firewall for AI agents

A firewall for AI agents: block prompt injection, guard tool calls, redact PII/secrets — a checkpoint between the agent you ship and the world.

pip install agentbastion

What it does

Three guards in one SDK: inbound injection block, tool-call guard, outbound PII/secret redaction
Tool-call guard is the differentiator — stops mass email/delete/refund/exfil
Drop-in SDK: data never leaves your box; optional LLM judge
Encoded attacks: bastion.decoded_payload decodes base64, hex, binary, rot13, leetspeak, reversed text and more; it ships in shadow mode and reports what it would block (58–74% of encoded corpus attacks, 0% benign false positives)
Honest benchmark: recall 1.0 / FPR 0.0 on the labeled corpus (public-set caveat in README)

from agentbastion import Bastion bastion = Bastion() result = bastion.check_input(user_text)

Version 0.14.0 · MIT + BSL-1.1 (gateway)