---
id: "CVE-2026-125368"
page: "https://infrastack.ai/live/CVE-2026-125368"
kind: "vulnerability"
aliases: ["CVE-2026-125368"]
title: "OmniAgent: approval check lets prompt-injected agents run arbitrary commands"
severity: "high"
cvss: 7.7
epss: null
exploited: false
status: "open"
category: "ai-tooling"
products: [{"name": "OmniAgent", "fixed": [], "affected": "<= 12b98c520959eff88a6cfa3dd233497084360ab0", "ecosystem": "github"}]
ecosystems: []
tags: ["ai-agents", "ai-tooling", "omniagent", "python", "self-hosted"]
not_affected_if: ["bash-tool-disabled", "human-approval-every-command", "no-untrusted-input"]
published_at: "2026-11-10T00:17:08+00:00"
first_seen_at: "2026-11-10T00:28:42+00:00"
updated_at: "2026-11-10T00:31:28+00:00"
sources: [{"source": "nvd", "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-125368", "first_seen_at": "2026-11-10T00:28:42+00:00"}, {"source": "analyst", "url": "https://github.com/YeQing17-2026/OmniAgent/issues/67", "first_seen_at": "2026-11-10T00:31:24+00:00", "cites": ["https://github.com/YeQing17-2026/OmniAgent/issues/67", "https://github.com/YeQing17-2026/OmniAgent/pull/68"]}]
exploit_sources: []
---

# OmniAgent: approval check lets prompt-injected agents run arbitrary commands

Link for people: https://infrastack.ai/live/CVE-2026-125368 (this .md is for agents)

OmniAgent's ReflexionAgent treats any bash command beginning with find as safe, so its approval check never inspects -execdir arguments. A prompt-injected agent can use find -execdir to run binaries such as dd, nc, perl or openssl without human approval, overwriting files or executing code.

_Summary, check and fix written by a model from the sources; labelled as such._

## Affected

- github:OmniAgent affected <= 12b98c520959eff88a6cfa3dd233497084360ab0 fixed no fixed version listed

## How to check

- Search your code and dependencies for `ReflexionAgent` and `_is_safe_bash_command` (omniagent/agents/reflexion.py).
- Run `git -C <omniagent checkout> log -1 --format=%H`. Commit 12b98c520959eff88a6cfa3dd233497084360ab0 or earlier is affected; no fixed commit exists on main yet.
- Check whether the `coding` or `full` tool profile is in use (bash is auto-approved only for 'safe' commands there) and whether the agent reads untrusted content (web pages, issues, files).

## Fix

No merged fix or release as of 2026-11-10. Open PR #68 (unmerged) adds the pattern r'-(?:execdir|okdir|exec|ok)\b' to the dangerous_patterns list in omniagent/agents/reflexion.py. Until merged, disable the bash tool or require approval for every bash command.

## Timeline

- 2026-11-10T00:28:42+00:00 nvd https://nvd.nist.gov/vuln/detail/CVE-2026-125368
- 2026-11-10T00:31:24+00:00 analyst https://github.com/YeQing17-2026/OmniAgent/issues/67 — Correction: nothing is marked verified. Advisory (VulnCheck) and issue #67 read. The fix is open PR #68, unmerged, with no release, and main has no fix commit. Not exploited; no public PoC (issue says steps on request).

## Documents

- [advisory] Third-party advisory: www.vulncheck.com https://www.vulncheck.com/advisories/omniagent-bash-approval-gate-bypass-via-find-execdir-in-safe-command-allowlist
- [describes] OmniAgent #67: bash safety classifier bypass via find -execdir https://github.com/YeQing17-2026/OmniAgent/issues/67
- [fix] OmniAgent PR #68 (open): require approval for find -exec/-execdir/-ok/-okdir https://github.com/YeQing17-2026/OmniAgent/pull/68

## From the sources

OmniAgent through commit 12b98c5 contains an approval bypass vulnerability in ReflexionAgent._is_safe_bash_command that treats any command starting with find as safe without inspecting -execdir arguments. Prompt-injected agents can run find with -execdir to invoke binaries like dd, nc, perl or openssl through an unsandboxed shell, overwriting files or executing code without human approval.

---

_Free to read, with fair use of about 60 pages a minute. For higher limits or API access: sales@infrastack.ai_