MAQPNA

Learn

Prompt injection in AI agents: constrain the tool calls.

You cannot reliably stop a model being persuaded. You can stop a persuaded agent from acting on it.

Updated 31 October 2026.

FAQ

Common questions

Can prompt injection be fully prevented?

No known method does it reliably. OWASP says it is unclear whether fool-proof prevention exists. That is why the practical goal for agents is to limit what a fooled agent can do.

What is the difference between direct and indirect prompt injection?

Direct injection comes from the user's own prompt. Indirect injection is hidden in content the model reads, such as web pages, documents, emails or tool results. Agents meet indirect injection whenever they read outside content.

What is taint tracking for AI agents?

Taint tracking marks a session once it has read untrusted content. Policy rules then deny, or require approval for, tools that could send data out. It works without knowing whether an injection actually happened.

Are prompt injection classifiers useless?

No. They are a useful signal, especially when a detection feeds a policy decision. They should not be the only control, because they are probabilistic and can fail open.

Assume the model gets fooled. Limit the damage.

Taint a session on a local MAQPNA and watch the gateway refuse the exfiltration call.

More install options
curl -fsSL https://maqpna.com/install.sh | sh