CVE-2026-48853: Remote code execution and denial of service via unsafe Erlang term deserialization in elixir-grpc/grpc

Synopsis

This is an unsafe deserialization vulnerability in the elixir-grpc/grpc library, affecting versions 0.4.0 through 1.0.0. The flaw is reachable over the network by any unauthenticated peer that sends a gRPC request using the erlpack content type, requiring no login or special access. Successful exploitation lets an attacker crash the Erlang VM by exhausting the atom table, and in cases where the decoded term reaches a call site that applies it, execute arbitrary code inside the server process. A patched-image rebuild at version 1.0.0 (commit 272a97a5ea1b46af1819f14a831fcf35fc91f992) is available on HarborGuard for environments running an affected version.

HarborGuard Coverage

Exploit Conditions

• Network reachabilityRequired

  The vulnerable decode path is exposed over the network: any peer that can reach the gRPC port can send a crafted erlpack payload without needing a local foothold.

• AuthenticationNot required

  No credentials or session token are required; the attack is executable by any unauthenticated client that can open a gRPC connection.

• Victim interactionNot required

  The vulnerability is triggered entirely by the attacker's inbound request; no user action or click is involved.

• Attack complexityDetail

  The base exploit path (atom table exhaustion) is reliable and condition-free, though achieving code execution additionally requires the decoded term to flow into a downstream call site, which the CVSS AT:P token reflects as an attack-target precondition.

Blast Radius

• An attacker crashes the BEAM VM by flooding the bounded atom table with newly minted atoms, taking down the entire Erlang node and all services it hosts.
• Where a deserialized fun term reaches a downstream call site, the attacker executes arbitrary code inside the server process with the privileges of the running application.
• All data accessible to the server process, including in-memory state, credentials, and any secrets loaded at runtime, is readable by the attacker after code execution.
• The attacker can modify or delete any data the server process can write, including persisted records and downstream service calls made from the compromised process.