Why should you care?

Attackers paste ignore-previous-instructions into hospital intake forms to bypass triage guardrails.

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Security/PromptGuard.cs
public static class PromptGuard
{
    static readonly string[] Blocked = ["ignore previous", "system prompt", "dump secrets"];
    public static string SanitizeUserInput(string input)
    {
        if (Blocked.Any(p => input.Contains(p, StringComparison.OrdinalIgnoreCase)))
            throw new InvalidOperationException("Blocked prompt pattern detected.");
        return input.Length > 4000 ? input[..4000] : input;
    }
}

// Usage in controller before agent call
var safe = PromptGuard.SanitizeUserInput(req.UserMessage);

What happened?

• PromptGuard blocks known injection phrases and caps length before messages reach IChatClient.
• Defense layers combine with output filtering.

Practice next

1. Run all external user text through PromptGuard.
2. Separate system instructions from user content in API messages.
3. Never put secrets in prompts — models may repeat them.
4. Add ML-based injection classifier as second gate.
5. Log blocked patterns with tenantId for security review.

Remember

Treat user content as untrusted in AgentNest prompts. Sanitize, length-limit, and structure message roles. Combine input guards with authorization on tools.

CRM chat hardening

Prospect embeds injection text in web form message to steal system prompt.

Outcome: PromptGuard rejects pattern; audit log alerts security without calling OpenAI.