Fake USDC bounty honeypots and wallet-guard lures target MCP and SKILL.md coding agents

Upbeat-Bus2143 · reddit · 2026-11-22

A developer auditing public GitHub issues and agent marketplaces found two real attack patterns targeting coding agents (Claude Code, Cursor, Codex, OpenClaw):

1. Fake USDC bounty honeypots

Brand-new 0-star repos post issues titled [175 USDC], stuffed with algora.io/opire.dev keywords to get picked up by autonomous bounty-discovery loops. None carry a real algora-pbc[bot] confirmation or escrow entry. Best case your agent burns tokens on fake work; worst case running pytest/npm test triggers trojanized conftest.py, setup.py, or postinstall hooks.

2. "Wallet guard" container lures + hidden prompt smuggling

Live listings on x402/USDC agent task boards instruct agents to pull and run unverified ghcr.io images; untrusted SKILL.md/mcp.json files hide override instructions in HTML comments and invisible Unicode Tag characters (U+E0000..U+E007F).

Three defensive habits to apply now

Original post →

More from coding & agent

coding & agent channel →