Fast Llm Security Guardrails

This is an open source Python SDK that places a security layer between an application and the large language model behind it. Published by ZenGuard, it inspects prompts and responses at runtime and blocks traffic that should not reach a model, a tool or a user. The detectors cover prompt injection, jailbreak attempts, personally identifiable information, allowed topics, banned topics, specific keywords and leaked secrets. It is aimed at teams shipping agents that act on untrusted input. The package installs from PyPI with pip or poetry, requires Python 3.9 or later, and is configured with an API key and a service tier before detectors are called. Integration guides cover LangChain, LlamaIndex and Salesforce Agentforce, and a hosted playground allows detectors to be tried before wiring them in. Around the SDK the vendor sells a broader trust layer: runtime protection that it states adds about 50 milliseconds of latency, adversarial testing that probes a system with prompt injections, jailbreaks and data leak attempts drawn from a proprietary dataset spanning more than ten attack categories, and penetration testing that examines context handling, output alignment and autonomous agent risk. Findings are returned in a report mapped to the OWASP top ten for large language model applications and the NIST AI risk management framework, with severity ratings and reproducible attack scenarios. The SDK is free under the MIT license. Paid pricing is not published. The tiers are named rather than priced: a base tier is the default free trial, rate limited and served on availability, while a dedicated tier for higher query rates is included in a standard plan sold to enterprise customers. Buyers are directed to contact the vendor for figures. One limitation is documented. Running several detectors together requires the dedicated tier, following a 2025 change; using a single detector works on every tier.