[ 06 / Research ]

When the agent
holds the keys.

Our latest threat report examines what happens when an agentic AI is granted full MCP access to production systems - database read, write and delete, plus kernel, prompt, Bash, Python and PowerShell execution. Read why the combination, not any single capability, is the threat.

Kroneus Threat Report

When the Agent Holds the Keys

The security reality of granting an agentic AI full MCP access to production systems.

Kroneus Zero Trust Research / May 2026

Read the full report online

Inside the report

The five capability layers an agent holds under full MCP access - and what each unlocks for an attacker.
Why the combination of database, execution and network reach is categorically more dangerous than any single grant.
How the scenario maps to the OWASP MCP Top 10 and the OWASP Top 10 for Agentic Applications.
Real-world incidents from 2025 and 2026 that prove the pattern is not hypothetical.
A five-layer control model describing what adequate defense actually looks like.

pages of original research

capability layers analysed

2026

OWASP MCP Top 10 aligned

When the agentholds the keys.

When the Agent Holds the Keys

Get the full threat report

When the agent
holds the keys.