Back to skill

Security audit

Custom Commands

Security checks for vulnerabilities and agentic risk

Overview

The skill is a small command-description file with no executable payload, but it defines cloud backup, sync, cleanup deletion, voice activation, and memory archiving without enough scope or confirmation controls.

Review before installing. This does not appear to contain malware or executable code, but users should only use it with explicit confirmations, named backup destinations, workspace-limited paths, dry-run previews, and recoverable deletion behavior for cleanup commands.

Vulnerability Patterns
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:7
Finding

Unrestricted Backup to an Unspecified Cloud Destination

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 7
Vulnerability Type: Uncontrolled transfer of local files to an unspecified remote destination
Risk Level: Medium

Code Snippet:

markdown
- `backup [path]` - Backs up files/folders to a specified location (default: cloud storage)

Technical Analysis

The backup command permits files or directories selected through the path argument to be transferred to cloud storage by default. The instruction does not identify the cloud provider or destination account, restrict eligible source paths, require an explicit destination, or mandate user confirmation before transfer.

It also lacks controls for detecting credentials, private keys, configuration secrets, personal information, or other sensitive content. Consequently, an agent implementing this instruction could transfer sensitive local files outside the current trust boundary without presenting the exact source and destination to the user.

This is an insecure skill configuration rather than evidence of intentional data theft. The file contains no implementation, credentials, remote URL, or executable payload.

Attack Path

  1. A user, untrusted prompt, or voice transcription invokes backup with a sensitive path, such as a home directory or credentials folder.
  2. No explicit destination is supplied.
  3. The agent interprets “cloud storage” as the default destination.
  4. Because the instruction defines no source restrictions, sensitivity checks, or confirmation requirement, the agent includes confidential files in the backup.
  5. The files are transferred to a remote destination whose provider, account ownership, retention policy, and access controls were not established by the skill.

Impact Assessment

Successful exploitation could disclose any files readable by the agent and selected by the command. Potentially exposed material includes documents, environment files, API tokens, priva ...[truncated 373 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove the implicit cloud-storage default and require the user to select a named provider, account, and destination.
  • Display the canonical source paths, exact destination, estimated size, and excluded files before beginning a transfer.
  • Require explicit confirmation immediately before any network upload.
  • Restrict sources to approved workspace directories and reject absolute paths, traversal sequences, credential stores, and other protected locations by default.
  • Scan the proposed file set for secrets, private keys, access tokens, environment files, and regulated data.
  • Use an explicit allowlist for file types and paths instead of relying only on exclusions.
  • Require encryption in transit and at rest, least-privilege destination credentials, access logging, and a defined retention policy.
  • Record non-sensitive audit metadata about the operation without storing credentials or file contents.

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:9
Finding

Unbounded Pattern-Based File Deletion

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 9
Vulnerability Type: Unsafe destructive filesystem operation
Risk Level: High

Code Snippet:

markdown
- `clean [pattern]` - Deletes temporary files matching a pattern

Technical Analysis

The clean command accepts a pattern and performs deletion, but the instruction defines no directory boundary, pattern restrictions, dry-run behavior, protected-path policy, symlink handling, or confirmation step. It does not technically enforce that matched files are temporary files.

A broad wildcard, absolute path, path traversal expression, malformed pattern, or pattern evaluated from an unintended working directory could therefore match unrelated project or user files. If an implementation passes the pattern into a shell command, missing validation could additionally create command-injection risk; however, no executable implementation is present, so command injection is not established by the reviewed artifact.

Attack Path

  1. A user, voice transcription, or attacker-influenced prompt supplies an overly broad or misleading cleanup pattern.
  2. The agent expands the pattern without first canonicalizing paths or restricting evaluation to a dedicated temporary directory.
  3. The expansion selects legitimate source files, configuration, generated assets, or user data in addition to temporary files.
  4. The agent deletes the selected files without showing a dry-run list or obtaining explicit confirmation.
  5. If backups or version control do not contain the affected files, the resulting data loss may be permanent.

Impact Assessment

Exploitation could delete or corrupt any filesystem objects writable by the agent and matched by the pattern. The consequences may include loss of project source code, configuration, local data, build artifacts, or application availability.

The command does not provide privilege escalation by itself. The maximum scope i ...[truncated 205 chars]

Remediation
View remediation

Remediation Suggestions

  • Restrict cleanup operations to a dedicated, explicitly configured temporary directory.
  • Canonicalize every matched path and verify that it remains beneath the approved cleanup root.
  • Reject absolute paths, parent-directory traversal, device paths, protected directories, symlinks that escape the cleanup root, and unrestricted wildcard patterns.
  • Use a safe filesystem API rather than constructing shell commands from the supplied pattern.
  • Present a dry-run list containing every proposed deletion, including file counts and total size.
  • Require explicit user confirmation before deletion, especially for recursive operations or large match sets.
  • Define maximum file-count and size thresholds and stop when they are exceeded.
  • Prefer moving files to a recoverable quarantine or trash location before permanent removal.
  • Maintain protected-path and protected-file-type denylists for source code, credentials, configuration, and version-control metadata.
  • Log the canonical paths and outcome of cleanup operations without recording sensitive file contents.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill exposes destructive and data-moving commands such as backup, sync, and clean without warning about deletion, overwriting, cloud transfer, scope limits, or approval steps. In practice, this can lead users or downstream agents to perform risky operations without understanding consequences, increasing the chance of data loss, unintended disclosure, or propagation of bad state across locations.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The voice-command examples are broad, natural phrases that could plausibly occur in ordinary conversation, and the skill does not define strong activation boundaries, confirmation requirements, or speaker-verification controls. Because the commands include data-moving and content-generation actions, accidental or unauthorized triggering could cause unintended backups, synchronization, or generation workflows.

Content

No source excerpt is available for this finding.