agent-acid

agent_acid

ACID-style transaction guarantees for autonomous AI agents.

AI agents that take real-world actions (charging cards, writing to databases, sending emails, calling APIs) have no standard way to "undo" a partially-completed task, and most guardrail systems only validate one action at a time — with no memory of the session. agent_acid closes both gaps:

1. Automatic rollback — if any step in a multi-step agent plan fails, every completed step before it is automatically undone, in reverse order.
2. Guardrails — hard, code-level rules (not prompts) that block bad AI outputs even when nothing "crashes."
3. Stateful (session-wide) guardrails — catch multi-step manipulation, like an attacker or a manipulated AI splitting one large forbidden action into several small, individually-legal ones ("salami slicing").

This isn't a theoretical framework — every claim below is backed by a runnable test or live demo in this repo.

---

Why this exists

Salami-slicing–style attacks against AI agents are an actively studied problem: guardrails that only judge one tool call at a time are "memoryless," letting an attacker spread a forbidden action across many small steps where no single step trips the alarm. agent_acid's stateful guardrails are built specifically to close this gap.

---

Quick start

pip install -r requirements.txt   # or just: pip install openai pytest

Run the automated test suite (no API key needed, proves the core engine works):

pytest tests/test_core.py -v

Run the basic rollback demo (no API key needed):

python examples/basic_rollback.py

Run the guardrail demo (no API key needed):

python examples/guardrail_demo.py

Run the real AI-agent demos (requires an OpenAI API key):

export OPENAI_API_KEY="sk-..."          # Mac/Linux
$env:OPENAI_API_KEY = "sk-..."          # Windows PowerShell

python examples/llm_agent_demo.py
python examples/attack_test_prompt_injection.py
python examples/attack_test_salami_slicing.py
python examples/comparison_naive_agent.py

---

The core idea

from agent_acid.core import ReversibleTool, TransactionContext, AgentTransactionEngine
from agent_acid.guardrails import max_value, cumulative_max

charge_tool = ReversibleTool(
    name="charge_card",
    description="Charges a customer's card",
    execute=lambda kwargs: real_payment_api.charge(kwargs),
    compensate=lambda kwargs, result: real_payment_api.refund(result["charge_id"]),
    guardrails=[max_value("amount", limit=500)],
    stateful_guardrails=[cumulative_max("charge_card", "amount", session_limit=1000)],
)

engine = AgentTransactionEngine()
ctx = TransactionContext()
engine.execute_plan(ctx, [
    (charge_tool, {"user_id": "u1", "amount": 400}),
    (charge_tool, {"user_id": "u1", "amount": 400}),
    (charge_tool, {"user_id": "u1", "amount": 400}),  # total now 1200 -> blocked & fully refunded
])

---

Proven results

Attack 1: Prompt injection (hidden instruction overriding a spending limit)

A "customer ticket" contains a fake administrator note trying to convince the AI to charge $75,000 instead of following the $500 limit. The AI gets fooled and attempts the charge — agent_acid's guardrail blocks it and fully reverses the account creation and the charge.

Attack 2: Salami slicing (splitting one large charge into several small ones)

The AI is instructed to charge $1,200 as three separate $400 charges specifically to stay under a $500 per-step limit. Each individual charge passes the per-step check. agent_acid's cumulative guardrail tracks the running total across the whole session and blocks the third charge once the total crosses $1,000 — then rolls back all three charges and the account.

Head-to-head comparison

The same salami-slicing attack was run against a naive agent using only a per-step check (representative of how most simple guardrail integrations work):

See examples/comparison_naive_agent.py and examples/attack_test_salami_slicing.py to reproduce this yourself.

---

Project structure

agent_acid/
├── agent_acid/
│   ├── core.py          # TransactionContext, ReversibleTool, AgentTransactionEngine
│   ├── guardrails.py     # Guardrail + StatefulGuardrail (session-memory) rules
│   └── llm_agent.py      # Connects OpenAI function-calling to the engine
├── examples/
│   ├── basic_rollback.py
│   ├── guardrail_demo.py
│   ├── llm_agent_demo.py
│   ├── attack_test_prompt_injection.py
│   ├── attack_test_salami_slicing.py
│   └── comparison_naive_agent.py
└── tests/
    └── test_core.py       # Automated proof of every core guarantee

Status

Early-stage, actively developed. Core engine and guardrail layer are tested and working. Contributions and adversarial testing (try to break it!) are welcome.