Check Point AI Guardrails
Check Point AI Guardrails screens user and external content going into LLMs and the resulting output, detecting threats and providing real time protection for a GenAI application and its users. It is the runtime layer of Check Point AI Agent Security and is also sold standalone for teams that want to embed the detection directly into their own AI applications. Detections cover prompt attacks (prompt injection, jailbreaks and manipulation in prompts, reference material, tool responses and tool descriptions), data leakage of sensitive information and PII, content violations (offensive, hateful, sexual, violent and vulgar content), malicious links outside an allowed domain list, off-policy agent behaviour via an Off-Task Action detector and a tool allow/deny list, plus custom controls. Integration is a per-interaction request to the Guard API against a project and policy, with a project mode that decides whether a flagged interaction is blocked or only logged. It runs as SaaS or as a self-hosted product. Trade-offs to weigh: it is proprietary with no published licence or source; screening is text only today, and the vendor states that defense against multi-modal threats in audio and images is coming soon; custom guardrails are listed as beta access on the SaaS tier; and the developer surface still carries Lakera branding (docs.lakera.ai, platform.lakera.ai, the LAKERA_GUARD_API_KEY variable) while the corporate site markets the umbrella product as AI Agent Security, so naming is inconsistent between vendor surfaces.
