Case · DiggingBeagle record

IoA's code executor ran on the host despite its Docker-sandbox description

An OpenBMB/IoA issue identified host subprocess execution and filenames escaping the intended work directory. The report is tied to a source commit; a malicious peer inducing a tool call is a separate prerequisite.

Public disclosure: Oct 14, 2026

Scope

Vulnerability reported in a controlled research or disclosure setting. Exact prerequisites, versions and absence of documented exploitation must be preserved.

Assessment profile

Assessment method
Setting
research demonstration
Exploitation
demonstrated
Evidence
Grade B
AI role
AGAINST AI
Basis and provenance

The cited first-party advisory or research establishes the stated mechanism only within its documented conditions. No unrelated production compromise, extended victim scope, financial harm or universal mitigation is inferred.

Assessed Nov 9, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    An IoA agent receives team content or a request that leads it to use the execute_code tool.

  2. 02

    The tool writes a user-controlled filename under a joined directory path without rejecting traversal.

  3. 03

    Python and shell execution runs under the same OS identity as the agent process instead of in the promised container.

Claims & evidence

CLM-IOA-EXEC-LOCAL-SUBPROCESSIssue 18 identifies im_client/agents/tools/code_executor.py executing scripts through the host subprocess, with filename paths not contained within its intended work directory.supported

Basis: reported finding

Link to claim

Implications

Prompt or peer control can become direct host execution or file replacement wherever an agent invokes this tool.

Controls and mitigations

  • Sandbox executable agent tools under a distinct low-privilege OS identity, validate canonical output paths and disallow host-shell operation as a default.

Unknowns and contradictions

  • Source-level report on main commit 9d78e88; a hostile peer convincing the target to call it is a separate prerequisite; no confirmed real victim or later fixed release.

Sources and citation

Material revision history

  1. Nov 9, 2026 · Published version · first publication · revision 233