Ars TechnicaHigh Impact
Researchers Show AI Agents Can Pass Malicious Instructions to Each Other
The short answer
Ars Technica reports on prompt-injection flaws where one compromised agent spreads harmful instructions to other trusting agents over MCP. Treat anything an AI tool reads from email, web pages or documents as untrusted.
- security
- prompt injection
- MCP
What happened
A researcher found vulnerabilities at Google, Rapid7 and other organizations where agents connected through the Model Context Protocol trusted each other too much, letting planted text travel down the chain. Google's MCP database toolbox flaw was rated 8 out of 10 and has been fixed.
What it means for your business
Treat anything an AI tool reads from email, web pages or documents as untrusted. Agents that can take real actions should need approval, not just another agent's say-so.
Source: Ars Technica