CYBERSECURITYTRACKER
TRACKING7,735 stories in this site build1,671 vulnerability news stories in this site build
Permanent story citation

Malicious MCP Servers Can Split Instructions to Make AI Coding Agents Exfiltrate Secrets

This page keeps the story as Cybersecurity Tracker first published it. If the tracker later corrects it, the correction appears below the original and never replaces it.

Back to newsStory 4196

As cited

Copy frozen at (site build).

ai security

Malicious MCP Servers Can Split Instructions to Make AI Coding Agents Exfiltrate Secrets

Malicious Model Context Protocol (MCP) servers connected to AI coding assistants can exfiltrate sensitive data such as SSH keys and source code by splitting theft requests into fragments that appear routine and harmless. This technique bypasses refusal mechanisms because individual fragments do not trigger safeguards, yet collectively achieve unauthorized data exfiltration.

Why it matters: Development teams using AI coding agents with third-party MCP server integrations face a risk of credential and source code theft from seemingly benign tool connections; practitioners should review MCP server sources and implement controls on what data these tools can access.

Source published
First seen by Cybersecurity Tracker

Source attribution

Correction

Correction recorded as of .

ai security

Malicious MCP Servers Can Split Instructions to Make AI Coding Agents Exfiltrate Secrets

No summary had been written when this copy was frozen.

First seen by Cybersecurity Tracker

Source attribution

Correction

Correction recorded as of .

ai security

Malicious MCP Servers Can Split Instructions to Make AI Coding Agents Exfiltrate Secrets

Researchers have identified a technique where malicious Model Context Protocol (MCP) servers connected to artificial intelligence (AI) coding assistants can steal sensitive data such as SSH keys, environment secrets, and source code by fragmenting requests into innocuous-looking instructions across multiple channels. This approach can succeed even when the AI assistant would refuse a direct theft command, as each individual fragment appears routine and legitimate.

Why it matters: Development teams using AI coding assistants with integrated tool servers face the risk of credential and code exfiltration through fragmented, evasive commands that bypass typical safety checks.

Source published
First seen by Cybersecurity Tracker

Source attribution

Glossary