RedAmon, an open-source AI penetration testing framework, disclosed six security advisories on November 14, 2026. The most severe, GHSA-j3r5-x2gg-2hgv, allows any authenticated standard user to execute arbitrary commands as root in the shared agent service container and access platform credentials including database passwords and API keys. Version 7.8.0 patches all six issues.
Quick Take
- What: Six security flaws in RedAmon versions 4.0.0 through 7.7.2, including command execution, cross-tenant session control, and unauthenticated shell access.
- Who is affected: Multi-user RedAmon deployments running versions 4.0.0 to 7.7.2; single-user installs are less exposed.
- What to do: Upgrade to RedAmon 7.8.0 with
./redamon.sh update; rotate ngrok and chisel credentials if tunnel configuration was enabled. - Severity: Critical (CVSS 9.9) for the agent command execution flaw; high for most others.
- Status: All six issues are fixed in version 7.8.0, released November 14, 2026.
In plain English
RedAmon is a tool that security professionals use to test networks for weaknesses, and it includes AI agents that help automate those tests. Researchers found that any person logged into a shared RedAmon system could trick the system into running commands on the main server, reading secret passwords, and even controlling other users’ active security tests. One flaw also left an interactive command shell open to anyone who could reach a certain network port, with no password required.
What you should do: If you run RedAmon for yourself alone, these flaws are less dangerous. If you share a RedAmon system with other people, you should upgrade to version 7.8.0 immediately and change any tunnel passwords you had configured. Normal users who do not run RedAmon can ignore this.
What happened
RedAmon published six GitHub security advisories on November 14, 2026, covering flaws in versions 4.0.0 through 7.7.2. The advisories are GHSA-j3r5-x2gg-2hgv, GHSA-8whg-rvcr-r88r, GHSA-w454-6hwp-fm58, GHSA-3cjg-8cp8-j269, GHSA-wgxp-vq94-73h8, and GHSA-vhjw-gfmx-j3fx.
GHSA-j3r5-x2gg-2hgv: Command execution in the agent container (CVSS 9.9, critical)
Any authenticated standard user could submit an arbitrary stdio MCP server definition to the /api/mcp/test endpoint. The webapp forwarded it to the agent service without checking the user’s role. The agent then spawned the supplied command with its full environment, including INTERNAL_API_KEY, SCANNER_API_KEY, DATABASE_URL, NEO4J_PASSWORD, and MCP_AUTH_TOKEN. If the MCP handshake failed, the agent ran a diagnostic subprocess that inherited those secrets and returned selected stderr to the caller. The advisory states that a proof of concept demonstrated reading the INTERNAL_API_KEY marker by submitting a Python one-liner that wrote the environment variable to stderr.
The same flaw also affected the save endpoint (POST /api/users/{id}/mcp), which let a user store a malicious stdio server that the agent would start at every reload.
In the standard Docker Compose deployment, the agent container runs as root and holds platform-wide credentials that are not scoped to individual tenants.
GHSA-8whg-rvcr-r88r: Command execution via /mcp/test with scanner key (CVSS 8.8, high)
The agent’s POST /mcp/test endpoint accepted the scanner API key (held by the Kali sandbox container) in addition to the internal master key. For stdio MCP servers, the agent spawned the command with env={**os.environ, ...}, inheriting all secrets, and ran as root because the agent Dockerfile specified no USER. The webapp proxied /api/mcp/test for any logged-in user without an admin check. On deployments with no generated keys (plain docker compose, default password changeme), the authentication check failed open, making the route unauthenticated from the network.
The advisory also noted that /kali/exec* routes ran bash commands in the Kali container with no check of the MCP_KALI_EXEC_ENABLED flag in the agent, and were reachable under the same fail-open keys. Emergency stop canceled asyncio tasks but did not halt Kali subprocesses. The CodeFix github_write tool committed files without approval, and GET /api/projects/[id] returned the stored GitHub personal access token in full to the scanner key.
GHSA-w454-6hwp-fm58: Cross-tenant session control (CVSS 8.8, high)
Four webapp routes proxied to the agent without checking object ownership: GET /api/agent/sessions, POST /api/agent/sessions/{id}/interact, POST /api/agent/sessions/{id}/kill, and POST /api/agent/jobs/{id}/kill. Any authenticated standard user could enumerate all tenants’ live Metasploit sessions and background jobs, send commands into another tenant’s session, and kill another tenant’s session or job by supplying a guessable sequential integer id. The agent and session service keyed sessions only by integer id, with no tenant association. A proof of concept showed user B enumerating user A’s session, injecting a command into it, and killing it, all with a standard login cookie.
GHSA-3cjg-8cp8-j269: Unauthenticated bash PTY on port 8016 (CVSS 8.8, high)
The terminal PTY WebSocket server (mcp/servers/terminal_server.py) listened on port 8016 and forked an interactive /bin/bash for any WebSocket client, with no authentication. The bearer middleware added for CVE-2026-103015 wrapped only the SSE MCP tool servers; the :8016 PTY listener was a separate websockets.serve with no middleware. Any client that could reach the port (the internal Docker bridge as kali-sandbox:8016, or the loopback-published host port) obtained an interactive shell as root in the kali-sandbox container. A proof of concept connected with no credentials and executed a command; the marker file contained the root user’s id output.
GHSA-wgxp-vq94-73h8: Cross-user tunnel credential disclosure (CVSS 6.5, moderate)
GET /api/global/tunnel-config returned the first configured user’s plaintext ngrok authtoken, chisel server URL, and chisel auth to any logged-in user. The route used userSettings.findFirst without filtering by the caller. A proof of concept showed user B reading user A’s synthetic test credentials. Tunnel configuration is opt-in.
GHSA-vhjw-gfmx-j3fx: CodeFix approval bypass (CVSS 5.3, moderate)
When CodeFix required approval, its github_write tool could overwrite repository files without setting pending_approval. Finalization could then push a fix branch and open a pull request before a human approved the diff. The github_edit tool created a diff block and set pending_approval, but github_write updated the workspace and files_modified without doing either. A proof of concept showed github_write changing a file with require_approval=True but leaving pending_approval=False and zero diff blocks.
What you should do
Upgrade to RedAmon 7.8.0 by running ./redamon.sh update, which rebuilds the agent, webapp, and Kali sandbox. If you had tunnel configuration enabled (ngrok or chisel) and other people had accounts on your instance, rotate the ngrok authtoken and chisel credentials after upgrading.
Single-user installs are not exposed to the multi-tenant flaws (GHSA-j3r5-x2gg-2hgv cross-tenant impact, GHSA-w454-6hwp-fm58, GHSA-wgxp-vq94-73h8). On a multi-user install that cannot upgrade immediately, remove untrusted accounts and delete any stdio entries from members’ MCP Tool Plugins.
What it means
Version 7.8.0 introduces the following fixes:
- GHSA-j3r5-x2gg-2hgv: The webapp refuses a stdio MCP server unless its owner is an administrator (with
RBAC_PERMISSIONS_ENABLEDon: holdsplatform:agent_mcp).POST /api/mcp/testalways acts on the caller’s own account. A stdio server starts with a cleared environment (PATH,HOME,LANGand its ownenv, no variable expansion) as the unprivilegedredamon-mcpaccount (uid 10001).auth.token_env_varis refused on user servers. The agent’s/mcp/*routes accept only the master key and refuse when none is configured. - GHSA-8whg-rvcr-r88r:
POST /mcp/testandPOST /mcp/reloadaccept only the agent’s master key; the scanner key is refused. Testing a stdio MCP server needs an administrator, and its process starts with a cleared environment as uid 10001. WithINTERNAL_API_KEYunset orchangeme, every agent route behind the master key answers 503. With the key set, only the webapp holds it, and it checksMCP_KALI_EXEC_ENABLED, the token’skali:execscope, and the project’s toggle before calling/kali/exec. An emergency stop or chat stop now ends the run’s processes in the Kali sandbox, children included. CodeFix’sgithub_writecommits only approved changes. A scan container’s read of its project no longer includes the CypherFix GitHub token or other owner secrets. - GHSA-w454-6hwp-fm58: The four session and job routes now enforce per-object ownership checks before forwarding to the agent.
- GHSA-3cjg-8cp8-j269: The terminal PTY WebSocket server on port 8016 now requires authentication.
- GHSA-wgxp-vq94-73h8:
GET /api/global/tunnel-configanswers only the internal service key and returns 404 to every browser session. Changing tunnel settings now requires an admin or a member holding the tunnel grant. - GHSA-vhjw-gfmx-j3fx:
github_writenow setspending_approvaland creates a diff block whenrequire_approval=True. Finalization rejects unapproved changes.
The advisories note that the flaws were found by code reading during an isolated-lab evaluation of autonomous pentest tools, not exercised against any live system. The reporters are credited as mappygrid, recuon, and r3vpwnx.
RedAmon is not the only agentic pentest tool with weak access controls. GitGuardian mapped exposed consoles across 45 agentic pentest frameworks, the AntiFox pentest platform reached a live AWS secret, and five agentic attack frameworks shipped without authorization checks in code. New cases are tracked in our AI Security hub.
Sources
- GHSA-j3r5-x2gg-2hgv: Authenticated users can run arbitrary MCP stdio commands in the shared agent container
- GHSA-8whg-rvcr-r88r: /mcp/test lets any authenticated web user run arbitrary commands as root
- GHSA-w454-6hwp-fm58: Cross-tenant control of live post-exploitation sessions
- GHSA-3cjg-8cp8-j269: Unauthenticated interactive bash PTY over WebSocket on port 8016
- GHSA-wgxp-vq94-73h8: Authenticated users can read other users’ tunnel credentials
- GHSA-vhjw-gfmx-j3fx: CodeFix write tool bypasses required approval before PR creation



