AIDEFEND in Action AIA-0151 Validated Research

VerCellBreak: Bypassing Vercel Sandbox SNI Egress Filtering via mTLS Client Certificate AIA Chasing

AIR Security’s VerCellBreak test starts with code inside Vercel Sandbox already able to read a secret. A direct connection to an unapproved host is blocked. An allowed external mutual TLS server can still send data on the sandbox’s behalf if its certificate verifier follows URLs from an untrusted client certificate. AIR’s service did so while looking for an intermediate certificate, carrying a test secret to a different host. The experiment shows why the outbound boundary must account for actions taken by an approved service after contact.

Data ExfiltrationEvasionRuntime IsolationAI InfrastructureAgentic AI
4 applicable AIDEFEND defenses
Source: VerCellBreak: How we escaped Vercel's hardened sandbox 
Authors: Matan Elgabsi, Or Nevo (AIR Security)
Original article: Oct 12, 2026

Threat Analysis

  • Two different connections matter. Vercel Sandbox uses Firecracker microVMs. In the tested configuration, its outbound policy used the TLS Server Name Indication (SNI) field to match the immediate TLS destination. That decision applies to the connection opened by the sandbox.
  • Mutual TLS can trigger work outside the sandbox. With server-authenticated TLS, the client checks the server certificate. In mutual TLS (mTLS), the server also checks the client certificate. The caIssuers entry in a certificate’s Authority Information Access (AIA) extension can help a verifier locate a missing issuer certificate. Some verifiers fetch missing issuers while building a chain.
  • What AIR actually showed. The experiment assumed code in the sandbox could read a secret. AIR connected to a permitted, controlled mTLS service that followed an AIA reference in a supplied client certificate. The verifier sent a separate request containing the test value to a second host. The first connection crossed the sandbox boundary under its allowlist, and the second was initiated outside it. AIR reports that Vercel treated this as out of scope because the external service issued the final request.
  • Who controls which part. The sandbox owner controls the guest’s route and what raw values are present. The operator of the allowed TLS service controls its verifier and its own outbound network. Vercel documents explicit egress policies and HTTP request mediation; supported plans can also keep certain outbound credentials outside the guest. Those controls must be configured for the actual workflow.

Applicable AIDEFEND Defenses (4)

AID-I-002.002
Secure External AI Service Connectivity
Very High
Route required AI HTTP requests through an organization-owned application gateway. Deny direct upstream connections and opaque tunnels. Any needed upstream mutual TLS is established by the gateway with its own managed client identity, so the external server cannot receive a sandbox-chosen client certificate on that route.
AID-I-004.007
Task-Bounded Context Segmentation & Secret Demotion
Very High
Keep the protected value in an external broker. Use a one-time scoped handle in the agent workflow; resolve it only for the approved operation outside the VM and return a bounded result. The raw protected value is absent from the sandbox and cannot be included in its client certificate.
AID-I-001.004
Sandbox Network Egress Restrictions
High
Bind every sandbox launch to an approved egress policy enforced outside the guest. Restrict the allowed destinations and protocols to what the workload requires; deny direct connections to external verifiers that do not have reviewed fetch behavior.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
High
If incident evidence shows that a long-lived key crossed this route, rotate or revoke the exact affected credential and confirm the old value fails at all relevant verifiers.

What Defenders Should Do Now

  • Review every allowed upstream that terminates or forwards requests. Prefer a mandatory application route with a controlled upstream identity and no direct guest access to other destinations.
  • Keep raw credentials outside guest files, environment variables and model context whenever an approved operation can be performed through a scoped broker.

1 additional consideration

Constrain issuer fetching at the external TLS service

The external service owns its certificate-chain builder and the network policy for lookups generated by the verifier.
Recommendation: Where the TLS stack supports it, pre-provision the approved intermediate certificates and disable automatic network issuer retrieval for untrusted client chains. Otherwise restrict that verifier’s DNS and outbound connections to approved CA endpoints. Verify both legitimate client authentication and refusal of external lookups in a controlled environment.

Conclusion

AIDEFEND ’s secret isolation and managed egress controls address the two decisions the sandbox operator owns. The TLS service operator can separately constrain lookups performed by its verifier.

Last updated: