crestapps-core-prompt-security

CrestApps.Core Prompt Security

A local, deterministic, regex-based layer for AI Profile chat: input validation, a security preamble and input delimiters, output filtering, rate limiting and audit logging. It is defense in depth, never a replacement for tool authorization.

Guidelines

• AddAISuite(...) (through AddCoreAIOrchestration()) registers the layer: 26 built-in rules, the scoring engine, IPromptSecurityService, IOutputSecurityFilter, IAIChatSecurityAuditService, IChatRateLimiter, IChatSessionStartRateLimiter and SecurityPromptOrchestrationHandler. There is no AddPromptSecurity(...) method.
• PromptSecurityOptions is not bound from configuration by the framework. Bind it from your own section or site settings.
• Profiles can only change anti-spam throttles (PromptSecurityProfileSettings). Injection detection, output filtering, preamble, delimiters, length and blocking threshold are global.
• Chat Interactions get input delimiters (EnableChatInteractionInputDelimiters, on by default) but not the preamble or injection blocking; the operator controls their prompt, model and tools.

Configure Site Policy

using CrestApps.Core.AI.Security;

builder.Services.Configure<PromptSecurityOptions>(builder.Configuration.GetSection("CrestApps:AI:PromptSecurity"));

builder.Services.PostConfigure<PromptSecurityOptions>(options =>
{
    options.MaxPromptLength = 6000;
    options.BlockingThreshold = PromptRiskLevel.High;
    options.CustomBlockedPatterns.Add(@"\binternal-canary-123\b");
    options.AnonymousMessageRateLimitTiers =
    [
        new ChatRateLimitTier { Limit = 5, Window = TimeSpan.FromSeconds(30) },
        new ChatRateLimitTier { Limit = 30, Window = TimeSpan.FromMinutes(5) },
        new ChatRateLimitTier { Limit = 150, Window = TimeSpan.FromHours(1) },
    ];
});