Lakera Guard
Type: full-code · Vendor: Lakera · Status: active · Status in practice: emerging
Lakera Guard wraps a real-time security firewall around a GenAI application, screening every user input, reference document, tool response, and model output so attacker-planted instructions cannot override the agent's objective.
Description. Lakera Guard is a runtime security platform that sits between an application and its LLM and inspects each interaction through a single Guard API call. It detects direct and indirect prompt injection, jailbreaks, content violations, and data leakage, and blocks compromised instructions reaching connected tools before agents act on them.
Agent loop shape. Sits inline on every user interaction and agent step, screening inputs and the model output before either is acted on.
Primary use cases
- Securing customer-facing GenAI assistants against prompt injection
- Protecting tool-using agents from compromised retrieved content
- PII leakage prevention in LLM I/O
Key concepts
- AI security firewall — A runtime layer between the app and the LLM that inspects every interaction for attacks before they are acted on.
- Guard API → input-output-guardrails — A single API call per interaction that screens inputs, reference content, tool responses, and outputs.
- Prompt Defense → prompt-injection-defense — The guardrail that detects direct and indirect prompt injection and jailbreaks.
- Data leakage prevention → pii-redaction — Detection, masking, and blocking of PII and confidential content in prompts and outputs.
Patterns this full-code implements —
- ★Prompt Injection Defense
Detects and blocks direct and indirect prompt injection and jailbreaks before they reach the model.
- ★★Input/Output Guardrails
A real-time API guardrail layer that screens both inputs and outputs of every LLM interaction.
- ★★PII Redaction
Data-leakage guardrails detect, mask, and block PII and confidential content in prompts and outputs.
- ★Multimodal Guardrails
Moderates offensive content across 100+ languages with ongoing image and audio support.
Anti-patterns avoided
- ✕Goal Hijacking
Scans every reachable input surface (user prompts, reference materials, tool responses, tool descriptions) for instructions that would override the agent's behavior.
- ✕AI-Targeted Comment Injection
The Prompt Defense guardrail scans retrieved and reference documents for hidden instructions embedded in the content the model reads.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: needs-verification