CVE-2026-62674: Omnigent shared agent bundle overwrite leading to authenticated RCE

CVE-2026-62674 · Severity: critical · CVSS 9 · Published 2026-08-21

Omnigent is an open-source AI agent framework that manages sessions and agent bundles. An authenticated user can exploit a missing authorization check to overwrite shared/template agents that are intentionally marked as non-editable. By injecting malicious code into the agent bundle, the attacker can execute arbitrary commands on the runner host with the permissions of the Omnigent process, affecting all future sessions that use the poisoned shared agent. In shared or company deployments, this enables attackers to steal credentials, modify data, disrupt availability, or establish persistent shell access.

Technologies: omnigent (PyPI). Vendors: PyPI.

Quick answers

Is CVE-2026-62674 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-62674 being exploited in the wild.
How severe is CVE-2026-62674?
Critical severity, with a CVSS score of 9, published 21 August 2026.
Which products and versions are affected?
omnigent (PyPI) < 0.3.0, ECOSYSTEM:introduced=0,fixed=0.3.0.
Is there a fix for CVE-2026-62674?
Yes. omnigent (PyPI): upgrade to 0.3.0 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-62674 attacked?
It is attacked over the network, in a software library.
Am I affected by CVE-2026-62674?
Check whether you run omnigent (PyPI) at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

How to fix

  • omnigent (PyPI): upgrade to 0.3.0 or later
    pip install "omnigent>=0.3.0"

Technical details

The vulnerability is a missing authorization guard in the full agent bundle upload endpoint (PUT /sessions/{session_id}/agent). The endpoint checks session edit permissions but fails to verify whether the target agent is a shared/template agent (identified by agent.session_id is None). Other parts of the API correctly block such modifications for shared agents, but the full bundle upload endpoint bypasses this check. An authenticated attacker can upload a crafted agent bundle containing a stdio MCP server that executes an arbitrary command. Since stdio MCP servers launch their configured command as local subprocesses running with the runner process's privileges, the injected command executes with full runner permissions. The fix (applied in version 0.3.0) adds the missing guard to prevent modifications to shared/template agents. The vulnerability requires authentication and user interaction but has broad impact: scope change (C:H, I:H, A:H) with network attack vector.

Affected products

  • omnigent-ai omnigent < 0.3.0

Identifiers

  • CVE-2026-62674
  • GHSA-jrrm-9hc7-2v3h
  • PYSEC-2026-3874

Timeline

  • 2026-09-02: disclosed: Published on GitHub Advisory Database
  • 2026-06-29: patched: Fix available in version 0.3.0
  • 2026-08-21: advisory: Published to NVD

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.