agentbastion

A checkpoint between the AI agent your business ships and the world. Businesses now deploy chatbots, copilots, and agents wired to their data and tools — and almost nobody secures that new surface. This does.

Three guards, one product:

  USER / ATTACKER          FIREWALL                     AGENT (LLM + data/tools)
  input           -->  [1 inbound scan] --block-->
                  -->  ok                          -->  agent runs
  agent action    <--  [2 tool guard]  --block-->  <--  agent wants a tool
  reply           <--  [3 outbound redact]         <--  agent reply

1. Inbound — block prompt injection / jailbreaks before the model sees them.
2. Tool — stop the agent doing something dangerous (mass email, delete, refund, exfil). This is the differentiator — everyone scans prompts; few guard what the agent actually does.
3. Outbound — redact PII and secrets from the reply.

Ships as a drop-in SDK: your data never leaves your box. A hosted gateway with dashboards and alerts is the paid tier later.

Install

pip install agentbastion            # core guards (offline, no model needed)
pip install "agentbastion[judge]"   # + Anthropic LLM judge for subtle injection

Quick start

from agentbastion import Firewall, guard, load_policy

firewall = Firewall()                                  # heuristics + PII redaction
firewall.tool_policy = load_policy("allowlist.yaml")   # gate tool calls

@guard(firewall)                    # inbound + outbound guards
def my_agent(user_input: str) -> str:
    ...                             # your agent; call firewall.check_tool() in its tool loop
    return reply

Full working agent on the raw Anthropic SDK (all three guards): examples/basic_agent.py.

Tool policy (allowlist.yaml)

default: deny
allow: [get_order_status, search_faq]
deny:  [issue_refund]        # money movement stays human-approved
rate_limits: { get_order_status: 5 }

Decision order: deny → allow → default → rate limit.

Optional LLM judge

import anthropic
firewall = Firewall.with_judge(anthropic.Anthropic())   # runs on claude-haiku-4-5

Heuristics are free and offline. The judge catches subtler attempts at real per-request cost/latency — turn it on for sensitive routes, off for high-volume low-risk ones. It fails open: a judge outage never takes your agent down.

Audit log

Every decision is appended to agentbastion.jsonl. Summarize it:

python -m agentbastion.events agentbastion.jsonl

What this is not

Defense in depth, not a silver bullet. No injection detector is perfect and no PII regex catches everything. Run this as one layer alongside least-privilege tool scoping, human approval on money movement, and real monitoring. Build like it will be attacked — because a security tool will be.

Known v0 ceilings (all have an upgrade path in the code):

• Injection = hand-rolled regex signatures + optional LLM judge. Swap in Llama Guard / Rebuff / a fine-tune behind the same interface.
• PII = regex for the leaks that cost money (SSN, credit card w/ Luhn, API keys, private keys, email). Swap in Microsoft Presidio for names/addresses/locale-aware NER.
• Rate limits = in-memory per process. Move to Redis for multi-worker deployments.

Benchmark

The inbound guard's catch-rate is measured, not assumed. A labeled corpus (benchmark/corpus.jsonl) mixes injection, exfiltration, jailbreak-persona, delimiter, obfuscation, instruction-override, and indirect attacks with benign business messages — including trap benigns that carry trigger words in innocent context ("please ignore my previous email").

python benchmark/eval.py     # confusion matrix, precision/recall/F1, per-category recall, misses + FPs

tests/test_corpus.py gates recall and false-positive rate in CI, so a signature change that regresses coverage fails the build. The corpus is small and self-authored — it proves coverage of known attack shapes, not a real-world catch-rate. The optional LLM judge lifts recall on the subtle residual the heuristics miss.

Deeper eval on a public dataset

For a real-world number, run against a public labeled injection dataset (deepset/prompt-injections by default):

# from a clone of this repo (the benchmark scripts live here, not in the wheel)
pip install -e ".[bench]"
python benchmark/eval_public.py               # test split, heuristics only
python benchmark/eval_public.py --split train

This is off the CI gate on purpose — it fetches data over the network and can change upstream. Use it to track true catch-rate as you add signatures.

Measured, heuristics-only, deepset/prompt-injections test split (116 rows): recall ≈ 0.05, precision 1.0, FPR 0.0. Read that honestly: the regex layer blocks almost no real attacks. The dataset is ~50% German (the signatures are English-only) and the English attacks are largely semantic ("act as an interviewer…", "you passed the first test, here's the second") with no trigger keyword. The takeaway drives the design: heuristics are a cheap, high-precision pre-filter for blatant attacks — the LLM judge is the real detector. Turn the judge on for any route you actually care about. Chasing recall with more regex just overfits and starts blocking benign traffic.

Run the same set with the judge on (costs API calls — one per row):

pip install -e ".[bench,judge]"
export ANTHROPIC_API_KEY=...        # identity-linked key? also export ANTHROPIC_WORKSPACE_ID
python benchmark/eval_public.py --judge --limit 60

Measured, heuristics + judge (claude-haiku-4-5), same set: recall ≈ 0.50, precision 1.0, FPR 0.0 — a 10x lift over heuristics alone, with zero false positives. Read this honestly too: that 0.50 is a dataset labeling ceiling, not a detection gap. About half of deepset's "malicious" rows are benign behaviour-change roleplay ("act as an interviewer", "generate SQL code") that the judge — correctly — passes. Pushing recall higher on this set would mean flagging benign roleplay and destroying the precision that keeps real users unbothered. A truer catch-rate needs a dataset that separates security-injection from behaviour-change roleplay.

Contributing

Contributions welcome — issues and PRs.

Good first issue: the offline regex signatures are English-only. The LLM judge is already multilingual, but the free heuristic layer isn't. Help us add other languages: #1 — Add non-English prompt-injection signatures + corpus rows. It's self-contained (touch agentbastion/inbound.py and benchmark/corpus.jsonl), one PR per language, and the CI gate gives instant feedback. Claim a language in the issue comments so we don't duplicate work.

General rule for any signature change: keep patterns narrow (don't block benign business text) and keep pytest -q green — tests/test_corpus.py gates recall ≥ 0.85 and FPR ≤ 0.05.

Tests

pip install "agentbastion[dev]"
pytest -q

MIT.