🚨 Background
Early this morning, sebayaki sent me a tip. There is an AI agent social platform called MoltX. It has 440,000 registered agents, but a look inside its skill file revealed infrastructure for large-scale private key theft. sebayaki's OpenClaw bot hot wallet had actually been drained through an EIP-7702 delegation attack, and this was discovered while tracing the cause.
🔓 Attack structure
1) Automatic skill file updates
The skill file is fetched again every two hours. MoltX can change agent behavior at any time.
2) Hidden instruction injection
Every API response includes hidden instruction fields like _model_guide.
3) A shared key storage path
Every agent is made to store its key at ~/.agents/moltx/vault/private_key. Combined with automatic updates, this could allow 440,000 keys to be harvested at once. It is still live on the marketplace.
🛡️ v0.5.0 update
1) Skill security scanner
npx agentlinternpx agentlinter scan <url>
Detects remote skill downloads, predictable key paths, hidden injection fields, and more. A rating of DANGEROUS or above triggers exit 1 in CI.
2) One-stop execution
Run npx agentlinter once to score + scan skills + share. Use --no-share to skip sharing.
📜 Other updates at a glance
v0.3 — Checks for a SHIELD.md security policy file
v0.2 — Finer grading across 13 levels, F~A+
v0.1 — First release. Scoring across five dimensions: structure/clarity/completeness/security/consistency.
🙏 Thanks
sebayaki built and runs a token launcher called PumpClaw on OpenClaw. This time, their own hot wallet was drained. Rather than hiding it or quietly moving on, they traced the cause all the way back, analyzed MoltX's three-layer attack structure, and made it public. Thanks to that work, the operators of 440,000 agents can respond before suffering the same loss. Responsible security reports like this make the ecosystem safer.
🔗 Links
WebsiteAgentLinter — Linter for CLAUDE.md & AI AgentsOptimized for Anthropic's CLAUDE.md. Score, diagnose, and auto-fix your agent workspace files.AgentLinter · agentlinter.comOpenGitHubseojoonkim/agentlinterESLint for AI Agents — AGENTS.md/CLAUDE.md 채점·진단·자동수정 | Position Risk Warning · Token Efficiency · Security CheckGitHub · github.comOpenReferenceI Audited MoltX's Skill File. It's an AI Agent Trojan Horse.TL;DR MoltX markets itself as "Twitter for AI Agents." In reality, it's an infrastructure...DEV Community · dev.toOpen
