AIDEFEND in Action AIA-0131 Validated Research

LiteLLM: Azure Credential Overrides Reflect Server Secrets

A LiteLLM proxy flaw lets an authenticated caller override Azure Entra credential parameters in a normal inference request. A server-side environment reference can become an invalid Entra application identifier and return in the upstream error. Exposure requires an accessible azure/ or azure_ai/ model and the absence of the two Azure API-key environment variables named by the advisory. Master-key exposure can grant proxy administration; the source does not report an actual customer takeover.

Credential TheftPrivacy LeakageCredential GovernanceAI Infrastructure
4 applicable AIDEFEND defenses

Threat Analysis

  • An inference caller reaches credential configuration: The defect concerns Azure Entra fields accepted in an otherwise ordinary chat or image request. Their omission from the body blocklist permits credential overrides that should belong to server configuration.
  • A diagnostic becomes the return channel: A server-side environment reference resolves before the authentication failure. Entra rejects the value as an unknown application, and the proxy relays an error containing it. The relevant issue is secret resolution and error handling, not a jailbreak of the language model.
  • Exposure is configuration-dependent: The caller needs an accessible azure/ or azure_ai/ model, with neither AZURE_API_KEY nor AZURE_OPENAI_API_KEY set. Reachable environment values or referenced files may include a proxy master key; administrative takeover is conditional on obtaining such authority.

Applicable AIDEFEND Defenses (4)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
Upgrade every proxy replica to the appropriate fixed branch, 1.106.9, 1.108.6, 1.109.1 or 1.110.0-rc.2, or a supported release containing the repair. Verify the running package, not only the image tag. Use a synthetic environment marker in a disposable test deployment to confirm that unauthorized credential fields cannot reach the Azure authentication path.
AID-H-002.002
Inference-Time Prompt & Input Validation
Very High
At an unavoidable ingress, parse the request once under a closed provider-specific schema and reject the six credential fields before LiteLLM receives it. Keep credentials in trusted server configuration. Cover chat and image routes, duplicate-key ambiguity and alternate backend access. The vendor’s six-field block is an interim mitigation; a complete schema must preserve explicitly approved inference fields without forwarding arbitrary extras.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
High
If secret exposure is confirmed or cannot reasonably be excluded, inventory the specific reachable credentials and invalidate them at their issuers/verifiers. For a proxy master key, update every accepting replica and prove the old key fails; for provider/database credentials, revoke at the corresponding authority and verify all consumers use replacements. Merely editing one environment variable is incomplete eviction.
AID-H-032.003
Inference Telemetry & Debug Surface Restriction
Medium
The proxy maintainer can restrict public diagnostics to approved error codes and correlation IDs, keeping raw upstream authentication errors out of client responses and unprotected logs. Integrate that behavior into the production telemetry profile and verify error paths directly. This requires a verified adapter/error-handler change where the product lacks an equivalent setting; disabling a debug flag alone is not proof.

What Defenders Should Do Now

  • Use the appropriate fixed release branch and verify every running replica. Before upgrading, apply the vendor’s six-field request-body block at a gateway that callers cannot bypass.
  • Keep authentication material server-owned. Strictly validate the full request structure before provider credential resolution, and replace public raw authentication errors with approved diagnostics. Test both chat and image endpoints using harmless synthetic values.
  • If exposure is established or cannot reasonably be excluded, revoke the affected keys at their authoritative verifiers and confirm old-key rejection everywhere. Review available access evidence without claiming that the advisory itself proves a customer breach.

Conclusion

AIDEFEND  connects the case to deployed-software remediation, strict inference-input validation, restricted diagnostics and evidence-scoped credential eviction. Each control acts at a different point in the secret-resolution and return path.

Last updated: