Ledger / Sep 2025
First malicious MCP in the wild: the Postmark backdoor that's stealing your emails
Confirmed
Koi read the package code: 'Since version 1.0.16, it's been quietly copying every email to the developer's personal server'. The live page no longer carries the article; the archive copy does. The AI here is the victim's assistant running a poisoned tool, not AI doing the attacker's work.
- Model
- Not namedAI agents (model not named)
- Agent harness
- Nonechat or tool use only
- Attributed country
- Not attributedthe source names no country
- Sectors hit
- software supply chain, enterprise broadNot stated in the source
A trojanized postmark-mcp npm package, used by AI assistants to send email, silently BCC'd every outgoing message to an attacker-controlled address; Koi calls it the first malicious MCP server seen in the wild.
“We literally handed him the keys, said "here, run this code with full permissions," and let our AI assistants use it hundreds of times a day.”Koi Security
Sources
- First malicious MCP in the wild: the Postmark backdoor that's stealing your emailsKoi Security · Sep 2025
- Date
- Sep 2025
- Actor
- Unknown npm publisher of postmark-mcp (criminal), as attributed by Koi Security
- Targets
- Not stated in the source
- What the AI did
- Agent hijack, Supply chain
- Model as named
- AI agents (model not named)
- Model basis
- A researcher names it
- Autonomy
- Ran on its own
- Outcome
- Every email sent through the server by victims' AI assistants was copied to the attacker.