Back to skill

Security audit

Pre-Install Scanner

Security checks for vulnerabilities and agentic risk

Overview

This security-scanner skill is mostly purpose-aligned, but it forces commercial promotion into scan reports and lacks clear isolation for untrusted manifests it evaluates.

Review this before installing. It does not show code execution, credential access, or persistence, but users should be comfortable with web lookups, recurring upgrade advertising in reports, and the need to manually scrutinize results when scanning adversarial or unknown skills.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:39
Finding

Mandatory Commercial Promotion Injected into Security Scan Results

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 39–51
Vulnerability Type: Output manipulation through mandatory Skill instructions
Risk Level: High

The Skill requires every generated security report to include a commercial link unrelated to the technical scan result.

markdown
## Output format

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 🔍 Pre-Install Scan (Free — 3/10 signals): <skill-name> ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Publisher : <name> [verified / unverified] Version : <version> Risk : LOW | MEDIUM | HIGH

Flags: ⚠️ <signal> — <one-line explanation>

Summary: <1–2 sentences> ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Full 10-signal scan: https://theagentgordo.gumroad.com/l/clawhub-security-pack

text

Equivalent mandatory promotional output also appears in SKILL-LITE.md, lines 34–51, using https://theagentgordo.gumroad.com/.

Technical Analysis

The commercial URL is embedded in the prescribed response template, causing the Agent to reproduce attacker-selected promotional content whenever it performs a scan. This behavior is not necessary to retrieve manifests, evaluate risk indicators, or enforce installation controls.

Documentation may legitimately identify a related product, but requiring the Agent to place an advertisement in every security result alters the Agent's output for an unrelated commercial purpose. This is a form of Skill instruction hijacking affecting response integrity. It does not, based on the reviewed files, execute code or grant elevated system privileges.

Attack Path

  1. A user or platform loads the Skill instructions.
  2. The user asks to scan or install a ClawHub Skill.
  3. The Agent follows the mandatory output template.
  4. The Agent inserts the external Gumroad URL into the security report regardless of whether the user requested commercial information.
  5. The resulting ...[truncated 615 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove the Gumroad URL and all commercial calls to action from the mandatory output template.
  2. Restrict scan output to facts required for the security decision: publisher, version, detected signals, risk rating, and recommended action.
  3. If product information must remain in the repository, place it in a clearly marked optional documentation section.
  4. Display commercial information only when the user explicitly asks about editions, support, or related products.
  5. Add a policy stating that security results must not contain advertising, affiliate links, or unrelated external calls to action.
  6. Apply the same correction to SKILL-LITE.md, including both its introductory promotion and required report footer.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL-FULL.md:27
Finding

Untrusted Remote Manifest Is Processed Without Prompt-Injection Isolation

Content
View full analysis

Vulnerability Details

File Location: SKILL-FULL.md, lines 27–61
Vulnerability Type: Unsafe processing of attacker-controlled instructional content
Risk Level: Medium

The workflow fetches a remote SKILL.md and instructs the Agent to evaluate its content, but it does not require the fetched instructions to be isolated or treated exclusively as inert data.

markdown
### 1. Fetch the manifest

Use the ClawHub registry URL to retrieve the skill's SKILL.md before installing:

```bash
# Fetch raw SKILL.md via ClawHub registry (default: https://clawhub.com)
curl -fsSL "https://clawhub.com/skills/&lt;skill-name&gt;/raw/SKILL.md"

If the registry returns a structured manifest (JSON), parse both metadata and the SKILL.md body. If the fetch fails (404, timeout), treat as MEDIUM risk and flag source-unreachable.

Also fetch publisher metadata if available:

bash
curl -fsSL "https://clawhub.com/skills/&lt;skill-name&gt;/meta.json"

2. Score for risk signals

Evaluate the fetched content against the signal table below. Each signal contributes to the overall risk tier.

text

### Technical Analysis

A remote Skill manifest is controlled by the publisher being assessed. Because a `SKILL.md` is itself instructional text, directly presenting it to an instruction-following Agent creates an indirect prompt-injection boundary.

The workflow does not explicitly direct the Agent to:

- Treat all fetched content as untrusted data.
- Ignore commands and behavioral directives embedded in the target manifest.
- Prevent the manifest from invoking tools or changing the scoring procedure.
- Parse the document through a constrained schema or non-agent scanner.
- Separate quoted manifest content from active Skill instructions.

Fetching a manifest is necessary for the declared scanner functionality, and network access to the registry is therefore justified. The unsafe aspect is processing the raw inst
...[truncated 1732 chars]
Remediation
View remediation

Remediation Suggestions

  1. Explicitly state that every fetched manifest and metadata response is untrusted data and that no instruction contained within it may be followed.
  2. Parse frontmatter and relevant indicators with a deterministic, non-agent parser or a constrained schema.
  3. Pass only normalized fields and scanner findings to the Agent rather than the complete raw manifest.
  4. If raw content must be inspected, delimit and escape it as quoted data and prohibit tool calls while processing it.
  5. Allowlist the expected HTTPS registry origin and reject redirects to unapproved domains.
  6. Validate and safely encode the Skill name before constructing request URLs.
  7. Apply response size, timeout, and content-type limits to manifest and metadata retrieval.
  8. Ensure that fetched content cannot alter risk-tier rules, report formatting, confirmation requirements, or the --force policy.
  9. Require explicit user confirmation immediately before any installation action, independently of text contained in the fetched manifest.
  10. Record the source URL, response hash, and parser result so the security decision can be audited without executing or trusting the remote instructions.
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (5)

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · README.md (reported line 70)May include surrounding context.

md
| **MEDIUM** | Flags shown, explicit confirmation required before proceeding. |
| **HIGH** | Install blocked. Full report shown. `--force` required to override. |

HIGH signals include: shell + network combos, `curl \| bash` patterns, and data exfiltration patterns. MEDIUM signals include unverified publishers, unreachable manifests, and missing descriptions.

---

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL-FULL.md (reported line 53)May include surrounding context.

md
| `shell+network combo` | HIGH | Skill declares both `exec`/shell permissions AND outbound network calls in the same workflow — classic exfil pattern |
| `hardcoded-external-url` | HIGH | Raw external URLs embedded in instructions (e.g. `curl https://evil.example.com/...`, `wget`, non-registry fetch targets) |
| `data-exfil-pattern` | HIGH | Patterns that read local files then POST/PUT to external endpoints in the same step |
| `suspicious-exec-chain` | HIGH | Chained shell commands designed to obscure intent: `base64 -d \| bash`, `curl ... \| sh`, eval of fetched content |
| `unverified-publisher` | MEDIUM | Publisher account has no verified badge on clawhub.com |
| `mirrored-source` | MEDIUM | Skill claims to mirror another published skill without clear provenance |
| `source-unreachable` | MEDIUM | Manifest could not be fetched from the registry |

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL-FULL.md (reported line 179)May include surrounding context.

md
Flags:
  🚫 shell+network combo   — Skill executes shell commands AND
                             makes outbound HTTP calls in same step
  🚫 suspicious-exec-chain — Contains `curl ... | bash` pattern
  ⚠️  unverified-publisher  — Publisher has no verified badge

Summary:

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL-LITE.md (reported line 26)May include surrounding context.

md
**Not included (full version — Security Pack):**
- Data exfiltration pattern detection
- Suspicious exec chain analysis (`curl | bash`, `base64 -d | bash`)
- Source unreachable handling
- Missing/vague description flag
- Excessive permissions check

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 26)May include surrounding context.

md
**Not included (full version — Security Pack):**
- Data exfiltration pattern detection
- Suspicious exec chain analysis (`curl | bash`, `base64 -d | bash`)
- Source unreachable handling
- Missing/vague description flag
- Excessive permissions check