Back to skill

Security audit

Remote Desktop Support

Security checks for vulnerabilities and agentic risk

Overview

The skill’s remote-support purpose is clear, but it can expose a live desktop and weaken local protections with failure modes that users should review carefully.

Only install this if you are comfortable giving temporary live desktop access and can supervise each session. Run preflight/status first, use view-only unless control is truly needed, avoid unlocking the session through the skill unless you explicitly accept that risk, and manually verify close/status afterward. Prefer a version that pins container image digests, fails closed when expiry or firewall checks are uncertain, keeps VNC encrypted or loopback-proven, and requires explicit confirmation before publishing the tunnel URL.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/remote_support.py:302
Finding

Unencrypted VNC May Be Exposed When the Firewall Is Inactive or Unavailable

Content
View full analysis
bool: if not firewall_cmd_available(): return False code, out, _ = run(["firewall-cmd", "--query-port", f"{VNC_PORT}/tcp"], timeout=10) return code == 0 and out.strip() == "yes" ``` ```python def enable_vnc(vnc_pass: str, mode: str) -> None: if vnc_port_allowed_by_firewalld(): raise RuntimeError(f"firewalld allows {VNC_PORT}/tcp; refusing to expose current-session VNC") set_vnc_encryption("['none']") ``` ### Technical Analysis The firewall check conflates two materially different conditions: 1. Firewalld is active and port 1024 is confirmed blocked. 2. Firewalld is unavailable, inactive, or cannot be queried. Both conditions return `False`. The caller interprets that result as permission to continue. It then explicitly configures GNOME VNC with encryption disabled. Although `preflight` reports firewall availability, `open` does not require proof that preflight succeeded. Consequently, direct invocation of `open` can start the VNC backend without an active firewall. The implementation only verifies that VNC is reachable through `127.0.0.1`; it does not prove that the backend is exclusively bound to loopback. This exceeds the minimum safe privileges required for the declared tunnel-based workflow because the desktop backend may become reachable through a host network interface rather than exclusively through the authenticated Guacamole tunnel. ### Attack Path 1. Firewalld is stopped, missing, or its query fails. 2. A user or automation invokes `remote_support.py open` directly. 3. `vnc_port_allowed_by_firewalld()` returns `False`. 4. `enable_vnc()` treats the result as evidence that the port is blocked. 5. GNOME ...[truncated 936 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/remote_support.py:499
Finding

Public Remote-Desktop Session Remains Open When TTL Scheduling Fails

Content
View full analysis
dict[str, Any]: sec = ttl_seconds(ttl) code, out, err = run([ "systemd-run", "--user", "--collect", f"--unit={OWNED_PREFIX}-expiry", f"--on-active={sec}", "/usr/bin/python3", str(Path(__file__).resolve()), "close" ], timeout=15) if code == 0: return {"method": "systemd-run", "seconds": sec, "detail": out} return {"method": "none", "seconds": sec, "error": err or out} ``` ```python expiry = schedule_expiry(args.ttl) state = { "status": "open", "created_at": time.strftime("%Y-%m-%dT%H:%M:%S%z"), "ttl": args.ttl, "mode": args.mode, "public_url": public_url, "local_url": f"http://127.0.0.1:{GUAC_LOCAL_PORT}/", "username": guac_user, "password": guac_pass, "containers": [GUACD_CONTAINER, GUAC_CONTAINER, TUNNEL_CONTAINER], "expiry": expiry, "previous_vnc_encryption": previous_vnc_encryption, "previous_lock_settings": previous_lock_settings, "unlock_info": unlock_info, } save_state(state) auth_check = verify_local_auth(state) one_click = guacamole_one_click_url(state) write_json(open_output(state, one_click, expiry, auth_check, one_click_only=args.one_click_only)) log("open", ttl=args.ttl, mode=args.mode, tunnel_started=True, expiry_method=expiry.get("method")) return 0 ``` ### Technical Analysis The declared safety model requires every open session to have a TTL and cleanup path. However, `schedule_expiry()` converts a failed `systemd-run` invocation into ordinary metadata with `method` set to `none`. `open_session()` does not validate th ...[truncated 1526 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
scripts/remote_support.py:40
Finding

Mutable Latest Container Images Allow Unreviewed Dependency Code to Execute

Content
View full analysis
Remediation
View remediation
`. - Optionally retain a human-readable version tag alongside the digest for maintainability, but use the digest as the enforcement mechanism. - Establish an explicit dependency-update process that reviews release notes, image provenance, vulnerabilities, permissions, and digest changes. - Verify signed image provenance with an appropriate mechanism such as Sigstore/Cosign where upstream support exists. - Apply restrictive container security options, including capability dropping, `no-new-privileges`, read-only root filesystems, and resource limits. - Avoid host networking for the tunnel container. Attach it to the dedicated Podman network and expose only the specific Guacamole endpoint it requires. - Minimize mounted credential lifetime and remove generated configuration files during `close`. ]]>

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/remote_support.py:563
Finding

Non-Atomic Session Check Allows Concurrent Open Operations to Corrupt Security State

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (9)

Lp1

High
Category
MCP Least Privilege
Confidence
97% confidence
Finding

The script launches and controls podman, loginctl, gsettings, systemctl, firewall-cmd, curl, and grdctl, which is substantial shell/process-execution capability not reflected in declared permissions. Undeclared command execution is dangerous because it permits system reconfiguration and remote-access setup beyond what a user may believe the skill can do.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
97% confidence
Finding

The script launches and controls podman, loginctl, gsettings, systemctl, firewall-cmd, curl, and grdctl, which is substantial shell/process-execution capability not reflected in declared permissions. Undeclared command execution is dangerous because it permits system reconfiguration and remote-access setup beyond what a user may believe the skill can do.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
97% confidence
Finding

The script launches and controls podman, loginctl, gsettings, systemctl, firewall-cmd, curl, and grdctl, which is substantial shell/process-execution capability not reflected in declared permissions. Undeclared command execution is dangerous because it permits system reconfiguration and remote-access setup beyond what a user may believe the skill can do.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The code can unlock the current GNOME session using loginctl when --unlock-current-session is supplied, expanding its power from 'share my desktop' to 'remove local session lock.' Unlocking a live user session materially weakens workstation security, especially if triggered remotely or without a prominent just-in-time consent flow, because it exposes all unlocked session data and applications.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/remote_support.py (reported line 73)May include surrounding context.

python
def run(cmd: list[str], timeout: int = 10) -> tuple[int, str, str]:
    try:
        p = subprocess.run(cmd, text=True, capture_output=True, timeout=timeout)
        return p.returncode, p.stdout.strip(), p.stderr.strip()
    except FileNotFoundError as e:
        return 127, "", str(e)

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill disables idle timeout, screen locking, and idle-activation settings for the desktop session, which broadens its impact beyond establishing a short-lived support link. These changes reduce endpoint security during the session and may persist if cleanup fails, leaving the workstation less protected after remote support ends.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill intentionally sets GNOME VNC encryption to ['none'], disabling transport protection for the live desktop session. Even if the design assumes localhost-only access, that assumption can fail through misconfiguration, container networking changes, local compromise, or unexpected port exposure, resulting in plaintext remote desktop traffic and weaker session protection.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

At the point where the session is opened, the script proceeds into logic that may unlock the current session and disable lock protections without a strong user-facing warning or confirmation in that execution path. For a remote-support tool, weak disclosure is dangerous because users may not realize they are authorizing actions that reduce local workstation protections, not merely generating a link.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The code starts a Cloudflare Quick Tunnel that exposes the Guacamole remote desktop service externally, yet there is no explicit runtime confirmation immediately before internet exposure. In context, this is particularly dangerous because the skill is creating live remote access to the current desktop session; a user should receive a clear, blocking warning before an externally reachable support URL is generated.

Content

No source excerpt is available for this finding.