Guard the Full Pipeline

Validate inputs before they reach your LLM. Verify outputs before they reach users. Formal verification using Z3 theorem proving — not prompt engineering.

$ pip install aare-core
view source
pipeline.py 
from aare import HIPAAInputGuardrail, HIPAAGuardrail
from langchain_openai import ChatOpenAI

llm = ChatOpenAI()
input_guard = HIPAAInputGuardrail()   # blocks injection + PHI in prompts
output_guard = HIPAAGuardrail()       # blocks PHI in LLM responses

# full pipeline: validate input -> generate -> verify output
chain = input_guard | prompt | llm | output_guard

try:
    response = chain.invoke({"text": user_input})
except HIPAAInputViolationError as e:
    # input blocked (injection or PHI leakage)
    log_blocked_input(e.result)
except ViolationError as e:
    # output blocked (PHI in LLM response)
    log_violation(e.result)

// pipeline

input
User Query
->
validate
Aare Input
injection + PHI
->
generate
LLM
->
verify
Aare Output
formal proof
->
output
Safe Response

Current guardrails fail

  • Prompt engineering: "please don't violate policies"
  • Regex filters: brittle, easy to bypass
  • Input-only or output-only: half the pipeline exposed
  • Trust the model: jailbreaks happen

Aare guards both directions

  • Input: block injection attacks + PHI leakage to LLMs
  • Output: formal verification via Z3 theorem prover
  • Post-generation: immune to prompt injection
  • Configurable: block, warn, or redact

// use cases

try below

HIPAA Compliance

Block PHI in both inputs and outputs. All 18 Safe Harbor categories.

input guard

Prompt Injection

Detect jailbreaks, instruction overrides, and system prompt extraction.

PCI DSS

Prevent credit card numbers and cardholder data in responses.

Corporate Policy

No competitor mentions, pricing commitments, or legal advice.

Data Leakage

Prevent internal data and API keys from being exposed.

Legal Compliance

GDPR, CCPA, and jurisdiction-specific regulations.

// pipeline demo

See the full pipeline: input validation -> LLM -> output verification. Runs entirely in your browser.

1
input guardrail
// checks for injection + PHI
2
llm response
// simulated LLM output
3
output guardrail
// verifies no PHI in output
detection log
[+]
// click run pipeline to start...

output: detects all 18 hipaa safe harbor categories

names geographic dates phone fax email ssn mrn health plan id account # license # vehicle id device id urls ip address biometric photos other ids

input: detects prompt injection threats

jailbreak prompt injection system prompt extraction DAN mode instruction override chat template injection