AIDEFEND in Action AIA-0138 Incident

NEAR AI: An Inference-Server Compromise Inside the Confidential Boundary

NEAR AI reported an inference-server compromise that reached internal legacy services. The case shows why confidential computing must be paired with secure application code, internal access boundaries and complete credential eviction: code executing inside the protected environment can still access the requests processed there.

Remote Code ExecutionCredential TheftRuntime IsolationAI Infrastructure
7 applicable AIDEFEND defenses

Threat Analysis

  • Confirmed and possible impact: The November 9 report describes SGLang remote code execution on one server, account-data retrieval for about 100 users and access to 18 legacy agents outside the TEE deployment. Two customers’ API keys appeared in retrieved logs. Live requests on the accessed server on November 8 could have been observed; stored Private Chat conversations and uploads were not shown to have been accessed.
  • The trust boundary: A trusted execution environment (TEE) protects computation against the host/operator threat model. It still executes application code. If that code is compromised, it can see data intentionally made available to it. Treating a valid enclave measurement as proof of safe application behavior would therefore miss this risk.
  • Keep the technical unknowns explicit: The notice does not publish the exploit input, CVE, fixed version or detailed route into legacy services. Those omissions prevent a defensible step-by-step exploit reconstruction; they do not prevent precise operator-owned remediation and containment recommendations.

Applicable AIDEFEND Defenses (7)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
High
The provider must rebuild and replace every affected inference runtime with the verified fix, including long-lived replicas and recovery images. Bind the fixed artifact to confidential-computing measurements where that platform requires them. NEAR reports a fleet fix but does not publish a CVE or version; a different SGLang advisory cannot supply the missing identity.
AID-I-002.001
Internal AI Network Segmentation
High
Place inference workers and legacy account/agent services in distinct internal trust zones. Enforce only documented service-to-service routes and test denied connections from the real worker identity. This can constrain the reported move into internal infrastructure without assuming which undisclosed port, credential or protocol the intruder used.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
High
Revoke exposed Cloud API keys and compromised administrative or long-lived credentials at their real issuers, then distribute replacements only to cleaned consumers. Verify old-key denial and legitimate replacement access. NEAR’s reported key revocation establishes a provider-owned operation; the framework’s AWS rotation script is not a NEAR API adapter.
AID-E-001.002
Issued Token, Authentication Session & Lease Revocation
High
Handle already-issued sessions and tokens separately from key rotation. Enumerate the incident scope at the provider’s issuers and gateways, invalidate each supported class, and verify rejection at the actual services. For self-contained tokens, issuer-session deletion alone is insufficient until relying parties deny them or verified expiry and containment close the window.
AID-I-005
Emergency "Kill-Switch" / AI System Halt
High
The provider should halt affected inference and legacy-service entry paths under an external incident decision while exposure is unresolved. Read back the effective halt across gateways, direct endpoints and recovery routes; separately stop active requests. The control limits additional exposure and must not be mistaken for removing the intruder’s process.
AID-E-002
AI Runtime Process, Lease & Pod Eviction
High
After preserving time-bounded evidence, the runtime owner must terminate the confirmed compromised inference process through a host- and instance-bound action, and prevent its supervisor from starting the same compromised runtime. Independently verify absence. Use the platform’s supported process or workload action; Linux pidfd requires an actual Linux host and appropriate operator access.
AID-R-004
Fleet Remediation Propagation & Technical Recurrence Prevention
Medium
Keep the provider’s validated fix in every applicable release and recovery baseline, with independent effective-version readback. Include a deployment test that rejects the old vulnerable artifact. This prevents a fixed production fleet from being repopulated by an old image during autoscaling or recovery.

What Defenders Should Do Now

  • Providers should repair and verify the full inference fleet, isolate internal data services, stop compromised execution and revoke both long-lived credentials and already-issued authentication state. Recovery must not restore the old artifact or redistribute new secrets into an unclean environment.
  • Customers should follow the provider’s account-specific notice and verify replacement credentials through the supported service. Assess live-request exposure separately from stored-data retrieval; do not assume the absence of one proves the absence of the other.
  • In an isolated representative environment, test that inference workers cannot reach unapproved legacy services, that revoked test credentials fail at the actual endpoints, and that old serving images cannot be redeployed. Require provider evidence for the undisclosed RCE regression.

Conclusion

AIDEFEND  places the defenses at the application, internal-network, execution and identity boundaries. Confidential computing remains useful for its stated threat model, while software remediation and verified eviction address code that has gained access inside that boundary.

Last updated: