Building BlinkGuard: A Pre-Sign AI Transaction Firewall & Trading Copilot on Solana

Ashlesha Anjarkar — Sat, 14 Nov 2026 09:22:32 +0000

Every month, Web3 users lose millions to blind-signing wallet drainers disguised as routine DeFi transactions. At the same time, autonomous AI agents executing on-chain swaps lack standardized preflight security guardrails.

Over the weekend for the MLH Solana Global Hackathon, I built BlinkGuard—a real-time, pre-sign AI transaction firewall and natural-language trading copilot on Solana.

⚡ The Problem We Solved

Blind Signing & Phishing Drainers: Standard wallet approval popups obscure complex contract logic, tricking users into approving unbounded token delegations or malicious escrow instructions.
Unsafe AI Agent Execution: Autonomous agents interacting with Solana DEXs are vulnerable to prompt injection and malicious swap routing without an independent verification layer.
DeFi UX Friction: Setting slippage thresholds, verifying token mints, and finding optimal liquidity routes across AMMs remains intimidating for everyday users.

💡 How BlinkGuard Works

BlinkGuard acts as a two-sided defensive layer on Solana:

For End Users: A conversational DeFi swap copilot powered by Gemini 2.5 Flash and Jupiter V6. You type plain English (e.g., "Swap 0.5 SOL to USDC"), and BlinkGuard simulates the balance changes, assigns a 0–100 Risk Score, and cryptographically attests the preflight check via Phantom Wallet (signMessage). If a drainer pattern or unbounded authority delegation is detected (Risk Score ≥ 80), the signature button locks automatically.
For Autonomous AI Agents: A standardized Solana Agent Skill endpoint (/api/skill) that any agent runtime can call as zero-trust middleware before signing a transaction.


text
User / Agent Prompt
       │
       ▼
[/api/parse] (Gemini 2.5 Flash Structured Schema)
       │
       ├─► Threat Detected (Risk Score ≥ 80 / Drainer Vector)
       │     └─► [LOCK] Signature button disabled & malicious vector flagged
       │
       └─► Intent Benign (Risk Score < 20)
             │
             ├─► [Jupiter v6] Optimized swap route fetched
             ├─► [Simulation] Expected in/out token balance diffs confirmed
             └─► [Wallet] Phantom signs cryptographic preflight attestation receipt

🛠️ Tech Stack

    Frontend & Framework: Next.js 15 (App Router), React, Tailwind CSS, Lucide Icons

    Blockchain & Wallet: @solana/web3.js, Solana Wallet Adapter (Phantom, Solflare)

    AI Engine: Google GenAI SDK (gemini-2.5-flash with strict JSON Schema evaluation)

    DEX Aggregation: Jupiter Aggregator API v6

    Deployment: Vercel

🤖 Solana Agent Skill Integration (/api/skill)

To make BlinkGuard composable for other builders, it exposes a machine-readable manifest and execution handler at https://blinkguard-uoq6-six.vercel.app/api/skill.

Any AI agent can audit an intent before signing using our TypeScript SDK (lib/solanaAgentSkill.ts):
TypeScript

import { blinkGuardAgentSkill } from "@/lib/solanaAgentSkill";

// Run preflight audit before initiating wallet signature
const result = await blinkGuardAgentSkill.execute({
  prompt: "Swap 0.5 SOL to USDC",
  userAddress: "YourSolanaPublicKeyHere",
});

if (!result.allowed) {
  console.error("Execution blocked by BlinkGuard:", result.reason);
}

🔗 Try It Out & Source Code

    Live dApp: https://blinkguard-uoq6-six.vercel.app

    Agent Skill Endpoint: https://blinkguard-uoq6-six.vercel.app/api/skill


  
    
      
      
        anjarkar25-max
       / 
        blinkguard
      
    
    
      
    
  
  
    

🛡️ BlinkGuard


Pre-Sign AI Transaction Firewall & Natural-Language Trading Copilot on Solana

Built for the Solana Global Hackathon (Major League Hacking)










⚡ The Problem




Blind Signing & Phishing Drainers: Users frequently authorize unbounded token delegations, malicious escrow instructions, or fake airdrop claims because standard wallet approval dialogs obscure transaction logic.

Autonomous Agent Security Risks: AI trading agents operating on Solana lack standardized, independent preflight execution guardrails, exposing them to prompt injections and compromised swap routes.

Complex DeFi Friction: Navigating decentralized exchanges, slippage thresholds, liquidity routes, and mint addresses presents a steep technical barrier for everyday users.






💡 Solution: What is BlinkGuard?


BlinkGuard delivers an active, two-sided defensive layer on Solana:






For End Users: A natural-language DeFi copilot that simulates transactions, scores threats (0–100), and dynamically disables signature actions upon detecting malicious drainer patterns.



For Autonomous AI Agents: A standardized Solana Agent Skill endpoint (/api/skill) and SDK (lib/solanaAgentSkill.ts…




  


  View on GitHub




I would love to hear your feedback on the architecture or ideas for expanding the preflight simulation rules! Drop a comment below if you're building with Solana or AI agents. 👇

How I Built Lazarus: Autonomous Legacy Resurrection with Kiro Specs, Custom MCP, and Differential Fuzzing

Ashlesha Anjarkar — Sat, 14 Nov 2026 07:23:42 +0000

Engineering teams sit on unmaintained legacy JavaScript modules they are afraid to touch because manual rewrites introduce silent edge-case regressions. For Kiroween 2026 (Frankenstein category), I built Lazarus — an autonomous legacy resurrection engine that stitches an isolated execution sandbox together with a 4-agent Kiro Crew and fast-check property-based differential fuzzing.

Live Operating Table: https://lazarus-yf0a.onrender.com
GitHub Repository & .kiro/ Artifacts: https://github.com/anjarkar25-max/lazarus
Social Blitz Post (#KiroweenHackathon26): https://x.com/ashlesha_25/status/2121502679369421082

Why Unit Tests Miss Legacy Regressions

Consider a classic ES5 32-bit integer truncation expression: (n / 2) | 0. A naive modernization replaces | 0 with Math.floor(n / 2). On positive integers, both match. On negative odd integers like -3, (-3 / 2) | 0 evaluates to -1, whereas Math.floor(-3 / 2) evaluates to -2.

How Lazarus Proves Behavioral Parity

Isolated Sandbox (src/sandbox/): Executes untrusted legacy code in a locked-down Node child process with strict memory, timeout, filesystem, and network restrictions.
Custom MCP Server (lazarus-sandbox): Exposes run_legacy_probe over stdio in .kiro/settings/mcp.json so agents safely probe legacy expressions.
4-Agent Kiro Crew (.kiro/agents/):
- Archaeologist probes boundary inputs to capture baseline runtime traces.
- Architect synthesizes observed behavior into formal Kiro specifications.
- Necromancer modernizes the code to ES2022+.
- Exorcist runs fast-check differential fuzzing (Object.is parity checks for -0 vs +0, NaN, and integer truncation). When Exorcist finds a divergence, it shrinks the input to a minimal counterexample and feeds it back to Necromancer to self-heal automatically.

Engineered with Kiro Spec-Driven Development

Every subsystem in Lazarus was designed using formal requirements, architecture designs, and implementation task lists inside .kiro/specs/, paired with repository-wide steering guidelines (.kiro/steering/) and an automated on-save Vitest hook (.kiro/hooks/run-tests-on-save.json) verifying 86 passing tests across 10 suites.

DEV Community: Ashlesha Anjarkar

Building BlinkGuard: A Pre-Sign AI Transaction Firewall & Trading Copilot on Solana

⚡ The Problem We Solved

💡 How BlinkGuard Works

anjarkar25-max / blinkguard

🛡️ BlinkGuard

⚡ The Problem

💡 Solution: What is BlinkGuard?

How I Built Lazarus: Autonomous Legacy Resurrection with Kiro Specs, Custom MCP, and Differential Fuzzing

Why Unit Tests Miss Legacy Regressions

How Lazarus Proves Behavioral Parity

Engineered with Kiro Spec-Driven Development