Phala incident combined unauthorized CVMs and evidence of sampled inference-secret misuse

Phala's preliminary incident report describes access through an internet-reachable legacy service into a loopback-trusting management path, allowing unauthorized confidential virtual machines on two hosts. The investigation separately addresses an underhardened GPU inference environment.

Investigators found that synthetic canary credentials included in selected prompts on October 21 and 22 were later used outside the platform. Phala rotated a KMS root key, took the affected inference environment offline and notified a specified CVM tenant. Confidential-computing marketing does not remove the need to secure host management and inference data paths. Canary reuse establishes exposure of some content or secrets, not continuous capture of every inference request.

The first-party report does not establish a TEE cryptographic break, a complete prompt-theft count, a common entry point for both affected surfaces or a known attacker identity. November 9 misuse of an already exposed credential is not necessarily renewed server access. Obtain the final root-cause report, isolate management routes, audit inference logs and evaluate whether sampled canary use can be tied to precise requests.

Evidence reconciliation and effect boundary. Two exposed surfaces must be kept apart: an internet-reachable legacy service could interact with loopback-trusting management to create unauthorized virtual machines, while a separate inference deployment lacked expected hardening. Later use of synthetic canary credentials supports exposure of some request content, but it does not by itself identify the extraction path or the total prompt population. The November 9 use of an already captured credential is also not evidence that attackers regained host control after containment.