/

CVE-2026-62674

Severity:

CRITICAL RISK

(9.0)

Code Injection (CWE-94) in Omnigent Shared Agent Bundles

CVE-2026-62674

Deferred

Vulnerability type

Category:

Code Injection

CWE:

CWE-94

Impact:

Remote Code Execution

Severity:

CRITICAL RISK

(9.0)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

Omnigent allows an authenticated user with session edit access to replace a shared or template agent without confirming that the agent is session-bound. By attaching an attacker-controlled stdio MCP server, the user may cause later sessions to execute commands with the permissions of the Omnigent runner.

Affected context

The vulnerability occurs in Omnigent deployments that use shared or template agents and expose the session agent replacement endpoint to authenticated users with edit permissions. The authorization check verifies session-level edit access but may not enforce ownership or binding requirements for the selected agent, allowing a shared agent bundle to be modified and reused by subsequent sessions.

Exploit preconditions

  • Attack surface: Authenticated HTTP API for session agent management
  • Required access: A valid account with edit permission for a target session
  • Trigger point: Assignment or replacement of an agent through the session endpoint, followed by use of that shared agent
  • Why exploitation is possible: Authorization does not sufficiently distinguish session-bound agents from shared or template agents whose agent.session_id is unset, allowing an executable MCP configuration to persist into later sessions

Why this happens

The underlying bug is an authorization and object-scope validation failure. The endpoint checks whether the caller can edit a session but does not verify that the selected agent is appropriately bound to that session before allowing replacement of a shared agent bundle. Because stdio MCP server configuration can launch external commands, attacker-controlled configuration is treated as executable behavior, mapping to CWE-94, Improper Control of Generation of Code.

Potential impact

  • The attacker may execute commands with the Omnigent runner process permissions.
  • The attacker could read workspace files, credentials, and other accessible data.
  • Internal services or connected environments may be probed or accessed.
  • Agent workflows and generated results may be manipulated or poisoned.
  • Runner availability could be disrupted.
  • Compromise could spread through shared agent configurations or connected tooling.

General mitigation guidance

Mitigation generally focuses on enforcing object-level authorization and

preventing shared configuration from becoming an untrusted execution path.

Common defensive patterns include:

  • Require explicit ownership and scope validation for every agent assignment.
  • Reject shared or template agents when their binding state is ambiguous.
  • Separate editable session configuration from globally reusable agent definitions.
  • Permit only approved MCP server types and commands through an allowlist.
  • Treat stdio command configuration as privileged executable content.
  • Run agent processes with least privilege, isolated workspaces, and restricted network access.
  • Avoid exposing credentials or sensitive files to runner processes.
  • Require review or controlled publication before shared agent bundles become available to later sessions.

How teams detect issues like this

Organizations can detect this class of issue through:

  • Static analysis and code review checks for authorization decisions that omit object ownership or scope.
  • Tests covering shared, template, unset, and session-bound agent identifiers.
  • Policy checks that flag user-controlled MCP server command or stdio configuration.
  • Runtime auditing of agent replacements, shared-bundle changes, and external process launches.
  • Alerts for unexpected commands, child processes, or network connections created by runner processes.
  • Automated security review tools that identify missing authorization predicates and unsafe command execution paths.

Sources

SourceLink
github.comgithub.com/omnigent-ai/omnigent/commit/25a22dc9e6da4648d23749f0a589e47e6aed991b
github.comgithub.com/omnigent-ai/omnigent/pull/1418
github.comgithub.com/omnigent-ai/omnigent/releases/tag/v0.3.0
github.comgithub.com/omnigent-ai/omnigent/security/advisories/GHSA-jrrm-9hc7-2v3h

Frequently asked

questions

It can be triggered through the Omnigent API when the deployment exposes the

affected session-management functionality and the attacker has the required

authenticated edit permission.

Not necessarily. Session-level edit access may be sufficient if the authorization

model does not separately restrict modification of shared or template agents.

Enforce object-level authorization and agent binding checks before accepting

replacements, and restrict MCP configurations to approved, non-user-controlled

commands.

Test every authorization path against shared and cross-session objects, treat

process-launch configuration as privileged, apply least privilege to runners,

and monitor changes to reusable agent definitions.

Find out what's already
exploitable in your codebase.