Stop AI Agent Incidents
Before Harm Accumulates

GuardRail monitors every agent output through API hooks or framework integrations. Each output is analyzed against six incident signatures simultaneously: PII exposure (regex + NER), cost anomaly (token count delta), hallucination markers (unverifiable claim patterns), prompt injection (adversarial instruction detection), latency spike (response time deviation), and semantic anomaly (output schema divergence). Detection runs in under 5 seconds from output generation.

Upon detection, GuardRail immediately quarantines the affected agent session. This means: suspend the session (prevent further output generation), block the violating output from reaching the user, revoke the agent's external tool permissions (prevents data exfiltration), and roll back to the last known good agent configuration. Containment is complete before anyone is paged — harm is limited in real-time.

After containment, the diagnostic engine reconstructs the incident timeline: what input triggered the incident? Was the agent's context poisoned (injected instructions, bad tool returns)? Is this a model degradation issue? Did a system prompt change cause the shift? The diagnosis is surfaced as a structured report with evidence — what changed, when, and why. This is the foundation for the remediation step.

Automated remediation launches in Phase 2. Based on the diagnosis, GuardRail will automatically apply fixes: add injection-resistant instructions to the system prompt, implement token budget guards, add output validation checks, or roll back to a previous agent version. All automated fixes include a human-in-the-loop safety check before execution.

After every incident, GuardRail generates a full post-mortem: timeline of events, root cause analysis, what was automated vs. what required human attention, and recommendations to prevent recurrence. Post-mortems are automatically shared with the relevant teams. The diagnosis engine learns from every incident across all customers — patterns become institutional knowledge that improves over time.