On this page
Exercising a real host-UID authorization boundary
Experimental study · 9 October 2026 · Draft pending review. A disposable Docker lab observation, not a security certification or GPU supervisor deployment.
Research question
Can a Linux service running under a different host principal from its client authorize that client, deny another principal that can connect to the socket, and block a third client through filesystem permissions?
Criteria frozen before acquisition
Protocol CUDA-05E-P01 was committed as 84f82a786d80981edd824a2d9c81f4bbfd105c71 before the trial. It independently tests three controls: actual service identity on the host, application authorization through SO_PEERCRED, and discretionary access control on the Unix socket.
The first setup attempt failed before any identity observation: chmod lacked sufficient authority after chown. A preregistered-protocol amendment was separately committed as fbd752d322a148f6d83279bef5cf04b4f9436374 before new acquisition. It reordered operations without granting additional capabilities. The aborted setup remains in the research history.
Observations from the completed trial
| Control or treatment | Observation | Limited interpretation |
|---|---|---|
| Service in a disposable Debian container | Its host UID differed from the permitted client's UID, corroborated by /proc and Docker configuration. |
Actual host-principal separation for this fixture. |
| Authorized client | Connection allowed. | Positive control. |
| Rejected client with group-level socket access | Connection succeeded, then SO_PEERCRED authorization denied it. |
Application UID rejection exercised. |
| Client without filesystem permission | Connection blocked before the server's accept. |
DAC denial independent of application authorization. |
| Authorized client after rejections | Connection allowed again. | Service continuity in this test. |
The service-owned directory used mode 0710; the socket used 0660. This is one completed experiment, without statistical repetitions. No CUDA/Vulkan workload was included.
What it does not establish
The Docker daemon remains privileged above the tested principals. Consequently, this does not qualify isolation against the host administrator, the real CUDA-05C Rust supervisor, physical GPU-loss recovery or production operations.
Full raw logs—including local process identity data—remain private. The public projection includes only bounded verdicts, a raw digest and pinned Git provenance. A checksum cannot independently disclose private evidence. External replication is not yet demonstrated.
Sanitized public observations · Experiment registry · Reusable experiment template
Discriminating next experiment
Run the actual CUDA-05C Rust supervisor under a dedicated service principal, with truly permitted and rejected clients, without risking the desktop session or resetting its GPU. Independently qualify IPC permissions, worker confinement, native provider integrity and recovery behavior.
Evidence class: single internal E4 observation. Production gate: DENIED. Status: draft; independent review pending.