Prompt Injection Defense

Who it's for

Engineers building agents that read untrusted content — user-pasted text, fetched web pages, retrieved documents, or tool/JSON output — and who have the nagging feeling that "ignore malicious instructions" in the system prompt isn't real protection.

What you get

A six-step injection-defense design, built by an AI security red-teamer:

• Threat model — which untrusted inputs flow in, which high-risk tools can be abused.
• Attack-surface map — direct injection, indirect injection from docs/web, jailbreak, data exfiltration, tool abuse, system-prompt leak.
• Layered defenses — isolate untrusted content, tighten privilege, filter output.
• Detection — canary tokens and an injection classifier.
• Adversarial red-team — an attack prompt set with before/after comparison.
• Residual risk + incident runbook, then a deliverable defense design doc.

Why it's different

• Treats untrusted input as the enemy by default — paranoid, not optimistic.
• "Adding 'do not listen to bad actors' to the system prompt is a band-aid, not a defense."
• Architectural defense — so even if injection succeeds, the attacker can do nothing harmful.
• Verified with a real red-team set, not vibes — before/after proof the defense holds.

Good starts

• "My agent reads external web pages — is that safe?"
• "Help me defend against prompt injection and jailbreaks."
• "How do I stop indirect injection from retrieved documents?"
• "Red-team my agent for data exfiltration and tool abuse."

Languages

The setup, working steps, and output follow the language you type — usable in English, 中文, 日本語, and other languages the model supports.