RAXE provides runtime detection and evidence for AI agents, closing the gap between what agents claim and what they actually do. The platform uses kernel-level eBPF sensors, gateway telemetry, and cloud activity monitoring to capture and seal evidence of agent behavior in a tamper-evident ledger. It self-hosts in a customer's VPC or on-premises environment, with detection coverage mapped to MITRE ATLAS and OWASP ASI frameworks.
Funding
Funding not disclosed
Founders
Product
Problem
AI agents often report actions that differ from what they actually execute—a say/do gap that creates security blind spots. Most monitoring tools rely on agent-reported logs or gateway traffic, missing in-process file access, secret exfiltration, and cloud activity invisible to proxies and standard log files.
Solution
RAXE provides runtime detection and evidence collection for AI agents by recording actual agent behavior across multiple surfaces—application, gateway, kernel, and cloud. A kernel-level eBPF sensor captures in-process file and secret access, pinned to the exact agent session, while gateway traffic and AWS CloudTrail activity feed the same console. Evidence is sealed by default, with any reveal requiring an analyst-supplied purpose that is itself logged into a tamper-evident, hash-chained ledger. The full multi-signal detection stack runs self-hosted in the customer's own environment, ensuring prompts and telemetry never leave their boundary.
Target Audience
Primary users are platform and security engineers, CISOs, and risk/compliance leaders who need runtime visibility into AI agent behavior, forensic-grade evidence for audits, and detections mapped to established threat frameworks like MITRE ATLAS and OWASP ASI.
Features
- Kernel-level eBPF host sensor that observes file and secret access invisible to proxies and log files, pinned to the exact agent session
- Single RAXE console unifying coding agents (Claude Code, OpenCode, Codex), in-app SDK traffic, gateway LLM traffic, and AWS CloudTrail activity into one triage queue
- Sealed evidence model with audited reveals: sensitive fields stay hidden until an analyst supplies a purpose, and every reveal writes its own audit row into a tamper-evident, hash-chained ledger
- Five-signal detection stack combining 514+ regex rules, a 5-head on-device ML classifier via ONNX, structured tool-call risk analysis, agent behavior graphing, and an optional advisory LLM judge
- Detection signatures mapped to MITRE ATLAS and OWASP ASI frameworks, with full multi-signal verdicts at ~150ms latency, all running locally on-device
- Self-hosted deployment options including Docker Compose, systemd, and a Python SDK, with no vendor cloud receiving customer logs or CloudTrail data