Refractal
Refractal provides security infrastructure for AI agents, sitting between agents and the tools, data, and people they reach to enforce policy at the moment each action happens. The platform hooks into agent actions before they run, applies deterministic policy rules, and sends findings to existing security stacks like SIEM, EDR, and NDR. It includes a statistical assurance layer that calibrates model probabilities and sets explicit error-risk bounds, with every decision written to a tamper-evident ledger.
- Cybersecurity
- Artificial Intelligence
- Regulatory & Compliance Technology
- Software Only
- AI Agents
- Developer Tools
- Enterprise Software
Funding
Founders
Founder details are not available yet.
Product
Problem
AI agents now act autonomously—browsing, querying, and executing transactions—in ways that are structurally vulnerable to prompt injection, jailbreaks, and data exfiltration. Each model carries different security policies, leaving companies without a unified layer to enforce safety and compliance across their agentic systems.
Solution
Refractal builds security infrastructure for high-assurance AI, positioning itself as a control plane for non-deterministic systems. It sits between agents and the tools, data, and users they interact with, enforcing policy at the point of action rather than after the fact. The platform operates through a four-stage pipeline: hooks trigger before key actions (tool calls, data reads, model swaps), rules are evaluated with minimal latency, events are blocked, flagged, or escalated based on severity, and all actions are recorded in an append-only, cryptographically signed audit ledger. This allows organizations to stress test agents before deployment, protect them in production, and maintain regulatory compliance through deterministic, testable, and versioned rules.
Target Audience
Primary customers are security, compliance, and ML engineering teams at enterprises deploying high-assurance AI agents that require strict governance, auditability, and protection against adversarial attacks.
Features
- Real-time hook system that captures identity, intent, and context before agent actions such as tool calls, data reads, and model swaps
- Policy engine that compiles rules from regulatory posture into deterministic, testable, and versioned evaluations with minimal latency
- Action routing based on severity, including blocking, flagging, or escalating critical events to on-call compliance teams
- Statistical assurance layer that calibrates model probabilities against held-out data and sets explicit error-risk bounds, with uncertain actions routed to human review
- Append-only, cryptographically signed audit ledger mapped to named controls for compliance reporting
- Integration with major observability and security stacks, including Langfuse, LangSmith, OpenTelemetry, CrowdStrike EDR, Splunk SIEM, Microsoft Sentinel, and Palo Alto Networks Cortex XSIAM
- Python SDK on PyPI and TypeScript SDK on npm, with new tenants starting in observe mode before enforcing policy
- Range run capability that attacks a replica of the customer's environment with adversarial multi-step campaigns before deployment