Advisory and proof-of-concept exploit for CVE-2026-123270 , an unauthenticated remote code execution vulnerability in the Redpanda Connect (formerly Benthos) streams-mode management API.
Coordinated through CERT/CC; tracked as CVE-2026-123270.
The recording runs the PoC end to end against the official v4.109.0 binary inside a throwaway container. Step [5] shows the blind path (command output written to a file on the host) and step [7] shows the in-band path (command output returned directly in the HTTP response body). The id payload runs as the unprivileged service account the container starts as ( uid=1000(demo) ), confirming code execution as the Redpanda Connect process with no authentication.
When Redpanda Connect runs in streams mode (for example redpanda-connect streams ), it exposes a management HTTP API on 0.0.0.0:4195 with authentication disabled by default. Any client that can reach port 4195 can create a stream with an unauthenticated POST /streams/{id} . Pipeline configs may contain code-executing processors such as command and subprocess , so an attacker can submit a pipeline whose processor runs an arbitrary command. Instantiating the stream executes attacker-controlled code on the host, with no authentication, credentials, or user interaction. The command output can be returned directly in the HTTP response body over the same management port, so execution is not limited to blind operation.
This is a secure-by-default failure (CWE-306, Missing Authentication for Critical Function) that escalates to code execution (CWE-94 / OS command execution).
It is a distinct issue from CVE-2026-82266, which affects the separate Redpanda broker Admin API on port 9644. See the advisory for the full comparison.
Vulnerable by default on all releases up to and including v4.110.0. From v4.111.0 (2026-09-25) onward it is not vulnerable by default, though the same RCE is still reachable there if the management API is enabled with the --bind-http flag. See the full advisory for the complete tested-version matrix (v4.39.0 through v4.117.0) and the details.
The PoC downloads and checksum-verifies the official Redpanda Connect release binary, starts it in default streams mode, confirms the management API is reachable without authentication, and then demonstrates two execution paths:
Blind execution: an unauthenticated POST /streams/{id} runs a command whose output is written to a file on the host.
In-band execution: an http_server input paired with a sync_response output returns the command output directly in the HTTP response body.
Run it against a local instance:
Run in a container (as recorded in the demo)
The demo above runs the target and the PoC together inside a throwaway container, as an unprivileged, non-personal service account. This isolates the vulnerable service from your host and is the recommended way to reproduce the exploit. It is also why the id output in the recording reads uid=1000(demo) rather than a real user account.
From the docker/ directory:
--rm discards the container after each run, and --hostname target keeps the host machine name out of any command output. The image downloads and checksum-verifies the official release binary at build time, and the PoC re-verifies the same digest at run time.
The build defaults to linux_arm64 . On an x86_64 host, override the two build args with the linux_amd64 asset and its published digest. This single swap covers both the build-time checksum and the runtime verification:
See docker/README.md for details.
CVE-2026-123270 was published on 2026-11-09 and fixed by the vendor. The captured proof-of-execution output and the reporter's email are redacted from this public copy.
Adrian Pastor (GitHub: @pagvac )
Released under the MIT License. See LICENSE .
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
