Loading vulnerability…
Authorization bypass in @ai-sdk/harness-codex relay
CVE-2026-64650 affects Vercel's @ai-sdk/harness-codex through version 1.0.28. The relay component used a process-path-based authorization fallback that trusted requests from any process whose command line contained the path of an approved helper script. In Linux environments, untrusted code already executing inside the sandbox could satisfy this check and invoke arbitrary host-exposed tools without a corresponding model-authorized tool-call event. This breaks the intended execution-layer authorization boundary between sandboxed code and host-side tool execution. Exposed operations could include sensitive actions such as secret retrieval, deployment actions, and cloud API operations. Version 1.0.29 removes the process-path authorization fallback and restricts relay acceptance to exact, short-lived, one-time authorizations tied to a bridge-observed model event.
What it means and what to do now: patch path, interim mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
Successful exploitation allows sandboxed untrusted code to bypass the intended model-mediated authorization flow and trigger host-exposed tools directly. Depending on which tools are exposed by the active harness session, this can result in unauthorized access to secrets, execution of deployment or administrative operations, use of cloud APIs, and other sensitive host-side actions. The vulnerability enables unauthorized tool execution rather than direct memory corruption or native code execution in the host process, but the downstream impact can be substantial if privileged tools are available.
If you can’t patch tonight, do this now.
Do not run the Codex harness on untrusted repositories or with untrusted dependencies until patched. When working with untrusted code, restrict host-exposed tools to low-sensitivity operations and avoid exposing tools capable of secret access, deployment, or cloud-administrative actions. Reducing or disabling host tool exposure during sandboxed execution lowers the impact if untrusted code is present.
Patch, then assume compromise.
Upgrade @ai-sdk/harness-codex to version 1.0.29 or later. The fix removes the process-command-line authorization fallback entirely and requires exact, short-lived, one-time authorization that matches the tool name and input to a bridge-observed model event.
No validated public exploits yet. Hidden: 1 rejected by Mallory's analysis as not a working exploit.
3 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.