AttoGuard

Real-time prompt injection defense for AI-controlled cryptocurrency wallets.

The Problem

AI-controlled cryptocurrency wallets are a growing category in DeFi. These wallets use large language models to interpret user intent and execute transactions autonomously. However, this creates a critical attack surface: prompt injection. Malicious actors can embed hidden instructions inside seemingly normal transaction requests, tricking the AI into sending funds to attacker-controlled addresses, approving malicious smart contracts, or leaking private keys.

Existing security solutions focused on smart contract auditing and transaction simulation, but none addressed the AI layer itself. There was no defense mechanism specifically designed to intercept and analyze the natural language inputs being sent to AI wallet agents before they reached the wallet logic.

The Solution

AttoGuard uses a proprietary, adaptive detection engine that scans wallet transactions for prompt injection attacks across 7 EVM networks in under 100ms. The system sits between the user input and the AI wallet agent, intercepting every transaction request before it reaches the wallet logic.

The engine learns from attacks it has already blocked, not from raw user input. Candidate signatures are extracted from confirmed attacks and only start contributing to detection once repeated sightings push their confidence past a fixed threshold. Supports 12 languages, detects encoding tricks and disguise attempts, and runs transaction simulations before execution.

Supported Networks

AttoGuard supports 7 EVM-compatible networks: Ethereum, Polygon, Arbitrum, Optimism, Base, Avalanche, and BNB Chain. Transaction simulation runs on the actual target network, so detection is accurate to current on-chain conditions.

Learning without poisoning

Adaptive detection is the obvious attack surface: feed a system enough crafted "attacks" and you can teach it to block legitimate commands. AttoGuard guards against that with a curated allowlist of normal wallet and DeFi phrasing - check balance, approve spend, bridge ETH, and hundreds more. Any candidate signature that overlaps legitimate usage is rejected before it can ever affect a real transaction, so the learning loop can strengthen defenses without being steered into false positives. Only high-confidence signatures are promoted to the shared threat feed across connected wallets.

Coverage is validated continuously, not asserted. A red-team suite exercises the production API every six hours across eight independent attack surfaces - a public injection corpus, crypto-native exploits, and novel injections generated fresh each cycle - and keeps a 90-day artifact trail of the results.

Results

AttoGuard currently detects 1,181,559+ distinct threat patterns across 12 languages with sub-100ms latency. The system supports real-time transaction simulation on 7 EVM networks and includes community-powered threat intelligence sharing across all connected wallets.