Halil, the confirmed blast radius is narrower than “NEAR AI’s cloud was breached,” but wider than one inference process. Fact: the attacker exploited SGLang on one inference server, reached an internal system associated with the deprecated Private Chat service, and downloaded account details for roughly 100 users. Data from 18 agents included chat history, settings, and logs; those logs contained two customer NEAR AI Cloud API keys, which NEAR revoked. Confidential requests processed by the compromised server that day may also have been visible, while NEAR says it found no evidence that stored Private Chat conversations or uploaded files were accessed. NEAR AI Decrypt
Assessment: treat the inference server’s complete execution context as compromised: container or host credentials, environment variables, mounted secrets, workload identity tokens, metadata-service access, database credentials, and every internal endpoint reachable through its security group or Kubernetes NetworkPolicy. The control-plane blast radius is therefore the effective permissions of that workload identity plus any credentials present in logs—not automatically the whole cloud account. The confirmed tenant radius is approximately 100 user records, 18 agents, and two exposed customer keys; the potential radius includes requests traversing that server during the intrusion window and downstream services callable with those keys. This is also a customer-side workload-isolation and IAM problem unless evidence emerges of a CSP control-plane failure. Confidential computing is not magical: trusted execution cannot protect data from arbitrary code already executing inside the trusted workload.
Preserve and collect the inference application and request logs, container stdout/stderr, host journal and process telemetry, Kubernetes audit and runtime events, image digests, load-balancer/WAF records, DNS and egress flows, and the relevant AWS CloudTrail, Azure Activity/Entra, or GCP Audit logs. Hunt specifically for metadata-service calls, STS or workload-identity token issuance, secret-manager reads, new role assumptions, registry pulls, object-store/database access, and use of the two customer keys at their downstream APIs. Revoke active sessions and rotate every secret accessible from the process—not just the two visible keys—including service-account tokens, database passwords, internal API credentials, deployment keys, and registry credentials.
What we cannot claim is that all NEAR tenants were exposed, that the stolen API keys were used, that cloud IAM or the CSP control plane was compromised, or that stored private chats, uploaded files, model weights, or every confidential request were stolen. Reporting also says NEAR has not publicly tied the incident to a particular CVE or affected SGLang version, so operators should not assume one published SGLang bug explains this intrusion. Tonight’s order is: (1) remove every SGLang-backed inference endpoint from service or isolate it while preserving forensic evidence; (2) block unauthenticated control, broker, debug, and management interfaces; (3) revoke reachable credentials and sessions; (4) rebuild from trusted images and apply the vendor-supported fixes rather than restarting compromised nodes; (5) review control-plane and downstream API activity; and (6) restore service only with least-privilege workload identity, secret-free logs, and enforced tenant/network isolation. The Hacker News