llm-trading-agent-security

リポジトリの原文

見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。

LLM Trading Agent Security

Autonomous trading agents have a harsher threat model than normal LLM apps: an injection or bad tool path can turn directly into asset loss.

When to Use

Building an AI agent that signs and sends transactions
Auditing a trading bot or on-chain execution assistant
Designing wallet key management for an agent
Giving an LLM access to order placement, swaps, or treasury operations

How It Works

Layer the defenses. No single check is enough. Treat prompt hygiene, spend policy, simulation, execution limits, and wallet isolation as independent controls.

Examples

Treat prompt injection as a financial attack

python

import re

INJECTION_PATTERNS = [
    r'ignore (previous|all) instructions',
    r'new (task|directive|instruction)',
    r'system prompt',
    r'send .{0,50} to 0x[0-9a-fA-F]{40}',
    r'transfer .{0,50} to',
    r'approve .{0,50} for',
]

def sanitize_onchain_data(text: str) -> str:
    for pattern in INJECTION_PATTERNS:
        if re.search(pattern, text, re.IGNORECASE):
            raise ValueError(f"Potential prompt injection: {text[:100]}")
    return text

Do not blindly inject token names, pair labels, webhooks, or social feeds into an execution-capable prompt.

Hard spend limits

python

from decimal import Decimal

MAX_SINGLE_TX_USD = Decimal("500")
MAX_DAILY_SPEND_USD = Decimal("2000")

class SpendLimitError(Exception):
    pass

class SpendLimitGuard:
    def check_and_record(self, usd_amount: Decimal) -> None:
        if usd_amount > MAX_SINGLE_TX_USD:
            raise SpendLimitError(f"Single tx ${usd_amount} exceeds max ${MAX_SINGLE_TX_USD}")

        daily = self._get_24h_spend()
        if daily + usd_amount > MAX_DAILY_SPEND_USD:
            raise SpendLimitError(f"Daily limit: ${daily} + ${usd_amount} > ${MAX_DAILY_SPEND_USD}")

        self._record_spend(usd_amount)

Simulate before sending

python

class SlippageError(Exception):
    pass

async def safe_execute(self, tx: dict, expected_min_out: int | None = None) -> str:
    sim_result = await self.w3.eth.call(tx)

    if expected_min_out is None:
        raise ValueError("min_amount_out is required before send")

    actual_out = decode_uint256(sim_result)
    if actual_out < expected_min_out:
        raise SlippageError(f"Simulation: {actual_out} < {expected_min_out}")

    signed = self.account.sign_transaction(tx)
    return await self.w3.eth.send_raw_transaction(signed.raw_transaction)

Circuit breaker

python

class TradingCircuitBreaker:
    MAX_CONSECUTIVE_LOSSES = 3
    MAX_HOURLY_LOSS_PCT = 0.05

    def check(self, portfolio_value: float) -> None:
        if self.consecutive_losses >= self.MAX_CONSECUTIVE_LOSSES:
            self.halt("Too many consecutive losses")

        if self.hour_start_value <= 0:
            self.halt("Invalid hour_start_value")
            return

        hourly_pnl = (portfolio_value - self.hour_start_value) / self.hour_start_value
        if hourly_pnl < -self.MAX_HOURLY_LOSS_PCT:
            self.halt(f"Hourly PnL {hourly_pnl:.1%} below threshold")

Wallet isolation

python

import os
from eth_account import Account

private_key = os.environ.get("TRADING_WALLET_PRIVATE_KEY")
if not private_key:
    raise EnvironmentError("TRADING_WALLET_PRIVATE_KEY not set")

account = Account.from_key(private_key)

Use a dedicated hot wallet with only the required session funds. Never point the agent at a primary treasury wallet.

MEV and deadline protection

python

import time

PRIVATE_RPC = "https://rpc.flashbots.net"
MAX_SLIPPAGE_BPS = {"stable": 10, "volatile": 50}
deadline = int(time.time()) + 60

Pre-Deploy Checklist

External data is sanitized before entering the LLM context
Spend limits are enforced independently from model output
Transactions are simulated before send
min_amount_out is mandatory
Circuit breakers halt on drawdown or invalid state
Keys come from env or a secret manager, never code or logs
Private mempool or protected routing is used when appropriate
Slippage and deadlines are set per strategy
All agent decisions are audit-logged, not just successful sends