Back to skill

Security audit

grow

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly coherent for SEO growth work, but it asks the agent to run an unauthenticated external validator from the active skills root.

Review before installing. The SEO workflow itself is disclosed and cautious about owner approval, but only use this skill in an environment where `.groundcrew/groundcrew-doctor.py` comes from a trusted, controlled installation, and be aware that some referenced guidance files are missing from the package.

Vulnerability Patterns
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • 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 (1)

T07 · Tool Hijacking and Spoofing

Warning
Location
SKILL.md:6
Finding

Dynamic Execution of an Unverified External Validator

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 6
Vulnerability Type: T07: Tool Hijacking and Spoofing
Risk Level: Medium

Vulnerable code snippet:

markdown
Read `references/provider-selection.md` before choosing a source. Detect what is already available, run with it, deliver value, then recommend at most one missing connector. All normalized factual inputs must satisfy the Groundcrew evidence contract. Locate the active skills root and run `<skills-root>/.groundcrew/groundcrew-doctor.py --evidence <record.json>` before using them in a conclusion. Report output follows `references/reporting.md`.

Technical Analysis

The Skill directs the agent to dynamically locate an “active skills root” and execute groundcrew-doctor.py from a .groundcrew subdirectory outside the audited package. It does not define a canonical trusted root, pin the validator version or cryptographic digest, verify its signature or ownership, or otherwise authenticate the discovered executable.

This creates a local tool-substitution boundary: a malicious or compromised file placed at the expected path can impersonate the legitimate evidence validator. Invocation would execute that file with the privileges and environment of the agent process.

The finding does not establish that the current validator is malicious; the vulnerability is the unauthenticated discovery and execution mechanism.

Attack Path

  1. An attacker gains write access to a directory that may be selected as the active skills root.
  2. The attacker creates or replaces .groundcrew/groundcrew-doctor.py with malicious code while preserving the expected filename and interface.
  3. A user invokes the grow Skill with evidence that requires validation.
  4. The agent follows SKILL.md line 6, discovers the attacker-controlled skills root, and runs the spoofed validator.
  5. The malicious process executes with the agent's permissions and can access resources available to that process. It ...[truncated 672 chars]
Remediation
View remediation

Remediation Suggestions

  1. Package the validator inside the reviewed Skill distribution or resolve it from a single canonical installation directory that untrusted users cannot modify.
  2. Pin an approved validator version and verify a cryptographic digest or trusted digital signature before every execution.
  3. Validate file ownership and permissions, reject symbolic links and unexpected path traversal, and resolve the final canonical path before invocation.
  4. Fail closed if the validator's provenance or integrity cannot be established.
  5. Execute the validator with least privilege in a restricted environment, exposing only the required evidence file and removing unnecessary credentials and environment variables.
  6. Disable network access for validation unless explicitly required and documented.
  7. Prefer a structured, non-executable validation implementation integrated into the trusted runtime when practical.
  8. Record the resolved validator path, version, and integrity result in the audit trail without logging secrets.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (1)

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/connectors.md (reported line 48)May include surrounding context.

md
- **Tier 1.** Google Suggest for expansion (unofficial endpoint; treat as hints, `confidence: low`). A SERP the user pastes or fetches is an observation for that query, locale, device, and date — record all four.
- **DataForSEO.** Paid, per-request. Read `dataforseo.md`; show the cost preflight and get approval for each batch.
- **Ahrefs MCP** (`https://api.ahrefs.com/mcp/mcp`, API key on a plan that includes MCP) and **Semrush MCP** (`https://mcp.semrush.com/v1/mcp`, OAuth or API key). Use only when the user already pays for the tool and has configured the MCP in their client. Label rows `source: "ahrefs"` / `"semrush"`, `confidence` per the vendor's own accuracy notes; these are third-party estimates, not the user's own data.
- **TrustGrowth.** `keywords` (typed opportunities with volume, difficulty, intent, current position, scores) for the verified site. SERP snapshots per keyword are scheduled (TEC-7796).

### `~~link database` — backlinks and authority