Questions about AI agent security.

Refractal is an AI security company for organisations that deploy AI agents. It post-trains compact security models on your environment and policies, checks every action your agents take before it runs, attacks a working replica of your environment to find weaknesses first, and records every decision in a tamper-evident ledger with a measured error rate.

AI agent security protects the actions AI agents take: the tool calls, file reads, shell commands, uploads and messages an agent carries out on someone's behalf. It covers prompt injection, including indirect injection hidden in documents, web pages and tool results, along with credential leaks, data exfiltration, destructive commands and agents acting beyond their permissions.

Refractal's models sit inline on every action an agent takes and allow it, block it or redact sensitive values before it runs. Because they judge the action itself, an injection that fools the agent's model still has to get its tool call past Refractal. They also read endpoint (EDR) and network (NDR) signals, so a compromised agent that behaves like a legitimate user still stands out.

With SPECTRE, Refractal's agent range. It builds a working replica of your environment, and adversarial agents attack it continuously with multi-step campaigns, from jailbreaks and indirect injection to tool-use abuse and data extraction, alongside your benign workflows. Your models are trained and tested there before anything touches production, and your real environment is never attacked.

Yes. Refractal works alongside your EDR and NDR: they see processes and traffic, and Refractal sees the action an agent was about to take. It exports its detections to your SIEM, including Google SecOps, and reads agent traces from OpenTelemetry, Langfuse and LangSmith's OpenTelemetry export.

Yes. Every decision Refractal makes is written to a tamper-evident, hash-chained ledger and mapped to the regulations you follow, including the EU AI Act, GDPR and DORA, with ISO 27001 and SOC 2 controls covered by an agent-controls pack. A statistical assurance layer gives each model a measured false-positive rate, and uncertain actions go to a person.

In your cloud or fully on-premises. Its security models are compact, built to match or exceed large general-purpose models at a fraction of the cost and latency, so they can run inside your own boundary.