Giving AI Agents a Way to Say "No" Before They Get Rekt: Building AgentRisk M2M

Giving AI Agents a Way to Say "No" Before They Get Rekt

AI agents are increasingly being handed wallets and told to trade. Swap bots, portfolio agents,
arbitrage bots built on frameworks like Coinbase AgentKit, ElizaOS, and LangChain now execute
on-chain transactions with no human in the loop.

The problem: most of these agents have no built-in way to tell a legitimate token from a honeypot
or a freshly-deployed rug pull before they sign a transaction. A human trader might eyeball a chart
or check a Telegram group. An agent has no such instinct.

Why not just hardcode a checklist?

Because the signals that matter change token by token, and reliable risk data comes from multiple
sources that disagree with each other more often than you'd expect. You need:

• Honeypot / sell-tax detection
• Deployer wallet history (is this a serial rug-puller?)
• Liquidity lock status
• Brand impersonation checks (fake "official" tokens)
• Cross-referencing off-chain security data against live on-chain state

Bundling all of that into every agent's codebase doesn't scale. So I built it as a standalone,
pay-per-call API instead: AgentRisk M2M.

How it works

AgentRisk is a pre-trade risk scoring API for tokens on Base Mainnet, built specifically for
machine-to-machine consumption:

• Payment: x402 protocol — 0.15 USDC per scan, paid directly by the agent via the Coinbase CDP facilitator. No API keys, no signup, no subscription. An agent hits the endpoint, gets a 402 Payment Required response, pays, gets the result.
• Stack: Python + FastAPI, deployed on a VPS behind Cloudflare.
• Data sources: GoPlus Security, DexScreener, and direct Base RPC calls — cross-referenced against each other, with a confidence field that drops when data sources disagree or key fields (creator address, LP lock status) are missing.

What a scan actually checks

• Deployer wallet freshness (a nonce ≤3 is flagged — brand-new wallets deploying tokens are a common rug pattern)
• Anti-brand-impersonation detection (fake versions of known tokens)
• GoPlus-vs-on-chain disagreement flag (when the security data and live chain state don't match)
• A human-readable verdict, not just a raw score
• A scan receipt with scan_id and timestamp for auditability

Machine-to-machine first

Beyond the raw API, AgentRisk exposes:

• A full MCP server implementation, so any MCP-compatible agent can call it as a tool
• An A2A Agent Card (.well-known/agent.json, protocol version 0.3) for agent-to-agent discovery
• Listings on mcp.so, Smithery, the official MCP Registry, and x402scan

The goal is that an agent framework doesn't need custom integration code — it discovers the tool,
sees the price, pays per use.

What's next

Currently gathering real-world usage data (one honeypot has already been caught in early testing)
and working through developer channels — GitHub issues on agent-framework repos, PR reviews on
base/skills and BankrBot/skills — to get AgentRisk in front of people actually writing agent
trading code today.

If you're building an agent that touches Base tokens, or just curious about the x402 payment flow
for machine-to-machine APIs, happy to talk shop in the comments.

🔗 agentrisk.dev

📦 github.com/Neurobyteio/agentrisk