Designed for infrastructure that handles sensitive workloads
The eBPF agent runs entirely inside your compute environment. We receive aggregated call-path telemetry, not model weights, not inference inputs, not output tokens. Your data classification boundary stays where it belongs.
Security by design, not by checklist
No model data leaves your boundary
The eBPF agent records call-path graphs and kernel execution durations. It does not read memory regions that contain weights, input tensors, or output tokens. The data type exfiltrated is timing metadata, not content.
No elevated privileges required
eBPF programs run in a sandboxed in-kernel VM with a fixed set of capabilities. zymtrace does not require root access, CAP_SYS_ADMIN, or kernel module installation. The kernel verifier validates the program before each load.
Data retention is configurable
Telemetry retention is 7 days on Trace, 30 days on Growth, and 90 days on Scale. On-premises aggregation mode is available on the Scale plan for air-gapped environments where call-path data cannot leave the datacenter.
SOC 2 Type II in progress
We are actively pursuing SOC 2 Type II certification. Our controls program covers access management, change management, and incident response. Contact us for a copy of our current security questionnaire or trust report draft.
Encrypted in transit and at rest
All telemetry is transmitted over TLS 1.3. Stored call-path data is encrypted at rest using AES-256. API tokens are stored as bcrypt hashes and are rotatable at any time from the dashboard.
Responsible disclosure program
We operate a responsible disclosure program and respond to valid reports within 5 business days. If you have found a vulnerability, email [email protected] with a proof of concept and we will coordinate a fix and disclosure timeline with you.
Exactly what zymtrace sends to our servers
Call-path graphs
Function names and call hierarchy from Python through to CUDA kernel names. No function arguments, no tensor contents.
Kernel execution duration
Per-kernel wall-clock time in microseconds. Duration metadata only, not kernel source code or compiled PTX.
GPU utilization metrics
SM occupancy percentage and memory bandwidth utilization. Aggregated per node, not per user request or prompt.
Model weights
The agent does not read memory regions mapped to model parameter tensors. Parameter data is not reachable via eBPF function tracing.
Inference inputs or outputs
Prompt text, input tokens, and generated outputs are not captured. The profiler operates at the kernel-dispatch layer, not the application layer.
End-user data
zymtrace has no visibility into who is sending requests to your inference endpoints or what those requests contain.
Review our full security documentation
Scale plan customers can request a security review, penetration test report, and a tailored data processing agreement.