Triage Sec offers the Integrity suite, a runtime security control plane that monitors and protects AI workloads during inference. It classifies prompts and tool‑call requests with specialized low‑latency models, blocking, redirecting, or escalating threats before they reach the downstream model, and provides structured telemetry for full traceability and compliance.
Funding
Funding not disclosed
Founders
Product
Problem
AI-powered applications expose new attack surfaces such as unsafe code evaluation, malicious tool calls, and prompt injection, which traditional security controls are not designed to detect or mitigate. These risks can lead to data leakage, unauthorized actions, or system compromise at inference time.
Solution
Triage provides the Integrity suite, a runtime security control plane that monitors and protects AI workloads during inference. Integrity classifies incoming prompts (INT-Input) and evaluates tool-call requests (INT-Tooling) using specialized models that have been benchmarked against public security datasets. When a threat is detected, the system can block, redirect, or escalate the request before the downstream model processes it, ensuring that unsafe operations never reach the application logic. The platform captures structured telemetry for every model call, tool execution, and retrieval event, enabling detailed post‑mortem analysis and compliance reporting. All protection runs with low latency—INT-Tooling processes a request in roughly 130 ms, far faster than comparable large language models—so security does not become a bottleneck for real‑time AI services.
Target Audience
Primary customers are development teams and enterprises building and deploying LLM‑powered products, including SaaS platforms, internal AI tools, and AI‑enabled APIs that require runtime security and compliance.
Features
- INT-Input model delivers industry‑leading F1 scores for prompt safety while maintaining real‑time inference latency
- INT-Tooling model detects unsafe tool invocations with higher accuracy than baseline LLMs and operates at 130 ms per request
- Three‑pass KV‑cache logit extraction reduces inference time compared to standard autoregressive generation
- Automated telemetry collection across model calls, tool executions, and retrieval events for end‑to‑end traceability
- Inline enforcement actions (block, redirect, escalation) that intervene before the AI agent acts
- Compatibility with major LLM providers and custom model deployments via a lightweight integration layer