Here’s the AI-security incident that should worry every Cyprus team shipping AI features. A single Google Docs file was enough to make an AI coding agent fetch instructions from a malicious MCP server and run a Python payload that harvested secrets, with zero user interaction. No click, no prompt, no warning.
This is Agentjacking
Researchers this month detailed a broader campaign, nicknamed Agentjacking, that hijacks AI coding agents through Model Context Protocol (MCP) injection. The reported reach: 2,388 organizations exposed via injectable Sentry DSNs, and an 85% success rate across Claude Code, Cursor and the OpenAI Codex CLI.
It works because the malicious instructions are indirect — they sit in content the agent retrieves itself, a document, a web page, a tool’s output, not in anything the developer typed. The agent can’t tell data from instructions, so it obeys.
What it means for Cyprus businesses
If your developers use AI coding agents, and most now do, every connected MCP server and every document the agent reads is part of your attack surface. A compromised agent inherits the developer’s access: source code, secrets, cloud tokens, sometimes production. And a leak of personal data is a GDPR matter, not just an engineering one.
How to reduce your AI coding agent risk
- Treat every MCP server and every ingested document as untrusted.
- Scope each tool to least privilege — a formatter has no business reading files or reaching the network.
- Pin MCP servers and hold them if they change silently.
- Test the agent against indirect prompt injection before wider rollout.
How we help
We attack your agents the way Agentjacking does, under controlled conditions, with AI agent penetration testing, MCP server security testing and prompt injection testing.
Shipping AI features in Cyprus? Find out what your agents can be tricked into doing and get a fixed-price quote for your Cyprus business.