Guardrails sit between your application and the model. Every input is classified before it reaches the agent, so unsafe prompts and tool calls are blocked before they run instead of patched after the damage.

What Guardrails detect

• Prompt injection attacks: attempts to manipulate the model through malicious instructions in user input or retrieved content
• Jailbreak attempts: techniques used to bypass the model's own safety mechanisms
• Unsafe tool calls: tool invocations outside your allowlist, stopped before they execute
• Sensitive data in transit: PII, PHI, and secrets redacted automatically

Detection runs on Superagent Guard, a family of small models purpose-trained for security classification. The detectors are trained classifiers, not system prompts or regex filters.

Structured decisions

Every check returns structured JSON: a pass or block classification, violation types, and CWE codes. Your system acts on the decision, and the trail stays auditable.

Open-weight and self-hosted

The same models that power the API are published as open weights on Hugging Face. Run them inside your own infrastructure: no data leaves your network, and typical latency is 50-100ms. See Open-weight models.

Prove it

Guardrails handle the safe part. To prove an agent holds up under attack, pair Guardrails with Red Team, which runs adversarial campaigns against your app, repo, agents, and models.

Next steps

• Open-weight models
• Red Team
• Findings