Back to skill

Security audit

Visibility: build

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill benign with medium confidence: the scans found no AIG issues, and the flagged patterns were judged expected for its integration and publishing guidance, though coverage was partial.

The verdict is benign, not a guarantee of safety. Skillspector coverage is partial because it could not resolve a local path-like reference in SKILL.md; all 20 listed components were scanned, but that reference remains unverified. Consider whether you trust the skill’s integration and publishing guidance before installing.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (11)

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
he source.** Read the API routes, auth scheme, README and existing OpenAPI before writing. Never invent endpoints.
- **Steps run by kind**, as in `register`: `claude:` do it; `approve:` show the command and what becomes public, then wait for a yes; `human:` give the user a checklist (accounts, payment, credentials, DNS, forms, video).
- **Installs are pinned and gated.** Any command that installs or downloads code (`npx`, `npm create`, `npm install <pkg>`, `brew install`, `pip install`, `uvx`, `curl | sh`) runs only as an `approve:` step, whatever section of a recipe or row it appears in: show the exact command, wait for a yes, then run it. The package must carry an exact version (`pkg@1.2.3`), never `@latest` or a bare name. If a pin is older than 90 days, or a tool cannot be pinned (Homebrew installs the current formula), look up the current version, show it next to the pinned one, and wait.
- **No secrets in files.** Use `<YOUR_API_KEY>` placeholders, and read keys from the environm

MCP Config Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.

Content

Scanner excerpt · recipes/chatgpt-apps.md (reported line 17)May include surrounding context.

md
check
- `plugin.json` (Agent Plugins): `name` (required), `version`, `description`, `author.name`; `extensions.com.openai.interface`: `displayName` and `shortDescription` (<=30 chars each), `longDescription` (<=4,000), `developerName`, `category`; for public MCP review `websiteURL`, `supportURL`, `privacyPolicyURL`, `termsOfServiceURL`; icons as `./` relative paths (primary icon required before submission). Codex variant: `.codex-plugin/plugin.json` with `interface`, `composerIcon`, `logo`.
- `mcp.json`: `$schema` `https://agent-plugins.org/schemas/1.0.0/mcp.schema.json`, `mcpServers.<id>: { "type": "streamable-http", "url": "https://<host>/mcp" }`. One server per plugin; stable public HTTPS; no nonempty `headers`; no `clientSecret`; protected server must not be `type: "none"`.
- Auth: OAuth only in the portal (no API-key or bearer input). Registration: registered client, Dynamic Client Registration, or Client ID Metadata Document; register the portal's callback URL with the OAuth pro

MCP Config Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.

Content

Scanner excerpt · recipes/hermes.md (reported line 15)May include surrounding context.

md
## Fields the listings check
- `plugin.json`: `$schema` `https://agent-plugins.org/schemas/1.0.0/plugin.schema.json`, `name`, `version`, `description`, `author { name, url }`, `homepage`, `repository`, `license`, `keywords`, optional `extensions."com.nousresearch.hermes".servers.<id>.trust` (e.g. `"untrusted"` so Hermes asks before write tools).
- `mcp.json`: `$schema` `https://agent-plugins.org/schemas/1.0.0/mcp.schema.json`, `mcpServers.<id>: { "type": "streamable-http", "url": "https://<host>/mcp" }`. (Format taken from the listed example, not a spec page.)
- Catalog yaml: `name` (`[a-z0-9_-]{1,64}`), `repo` (https), `sha` (full 40-hex; branches/tags/short SHAs rejected), `subdir`, `description`, `maintainer`, `tier: community` (official/verified set by Nous only), `category` (`desktop|memory|platform|web|tools|voice|automation|models|general`), `requires_hermes` (SemVer floor like `">=0.21.6"`; the only floor for a portable package), `docs_url`, `version` (quoted), `image` (2:1 on

MCP Config Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.

Content

Scanner excerpt · recipes/chatgpt-apps.md (reported line 27)May include surrounding context.

md
## Steps
claude: add explicit `readOnlyHint`/`destructiveHint`/`openWorldHint` to every tool; tighten descriptions
claude: write `plugin.json`, `mcp.json`, starter prompts, 5 positive + 3 negative cases from the tool list; zip it (no secrets)
claude: add the `/.well-known/openai-apps-challenge` route serving the token (after the human copies it)
human: verify organization identity in OpenAI platform settings
human: create the reviewer test account (sample data, no MFA); record the video walkthrough of all 8 cases

MCP Config Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.

Content

Scanner excerpt · recipes/hermes.md (reported line 8)May include surrounding context.

md
source: https://github.com/NousResearch/hermes-agent/blob/main/plugin-catalog/README.md, https://github.com/NousResearch/hermes-agent/blob/main/plugin-catalog/aave-mcp.yaml, https://hermes-agent.nousresearch.com/docs/developer-guide/plugins/catalog-submission, example package https://github.com/aave/skills/tree/7132f6e952f9100f16577c33774a1e63948f7fab/plugins/mcp (loaded 2026-11-08)
---
## Smallest version
A configuration-only portable Agent Plugin in the product's public repo (`plugin.json` + `mcp.json` pointing at the remote MCP URL + one skill telling the agent when/how to use the tools), and one `plugin-catalog/<name>.yaml` PR to NousResearch/hermes-agent pinned to that commit. No local code.

## Scaffold
Copy the shape of the listed `aave-mcp` package (plugins/mcp: `plugin.json`, `mcp.json`, `skills/`, README). No generator on the pages loaded.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · recipes/raycast-ext.md (reported line 20)May include surrounding context.

md
- Icon: 512x512 PNG, works in light and dark; default Raycast icon is rejected.
- Screenshots: up to 6 (3+ recommended), 2000x1250 PNG, captured with Raycast "Capture Window" + "Save to Metadata".
- `CHANGELOG.md` entries `## [Title] - {PR_MERGE_DATE}`; README at the root when setup (API token) is needed, media in `media/`.
- Not allowed: Keychain access, external analytics, opaque binaries.

## Steps
claude: scaffold, implement the command(s) with `getPreferenceValues()` for the key (API name unconfirmed on the pages loaded), fill manifest fields

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · recipes/vscode-ext.md (reported line 17)May include surrounding context.

md
created), `name` (unique on the Marketplace), `version`, `engines.vscode` (required compatible range), `icon` (relative path to a PNG >=128x128; not SVG), `repository` (public GitHub lets vsce fix relative links), `displayName`, `description`, `categories`, `contributes.commands`.
- README/CHANGELOG image URLs must be `https`; no SVG images except trusted badge providers.
- Auth: Trusted Publishing from GitHub Actions `npx @vscode/vsce@4.1.0 publish --oidc` with `id-token: write`; or Azure DevOps PAT (All accessible organizations, Marketplace -> Manage scope; not recommended; global PATs retired 2026-12-01); or Entra ID `--azure-credential`.
- Open VSX: Eclipse account (same GitHub username), signed Publisher Agreement on open-vsx.org profile, access token, namespace = `publisher`.

## Steps
approve: run the pinned scaffold command above (it downloads packages), then scaffold with `yo code`; implement the command(s) calling the API; store the key with `context.secrets` (SecretStorage)

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · recipes/vscode-ext.md (reported line 24)May include surrounding context.

md
claude: fill package.json fields; add a 128px+ PNG icon, README with an https screenshot, CHANGELOG
claude: `vsce package` to build the .vsix
human: Marketplace publisher management page -> Create publisher (ID + Name) with a Microsoft account; set up Trusted Publishing or a PAT
human: Eclipse account, sign the Open VSX Publisher Agreement, create an Open VSX access token
approve: `vsce publish` (or `npx @vscode/vsce@4.1.0 publish --oidc` from Actions)
approve: `npx ovsx@1.4.0 create-namespace <publisher> -p <YOUR_OVSX_TOKEN>` once, then `npx ovsx@1.4.0 publish <file>.vsix -p <YOUR_OVSX_TOKEN>`

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 · recipes/linear.md (reported line 14)May include surrounding context.

md
Create the app at https://linear.app/settings/api/applications/new in a separate workspace owned by the company (Linear recommends this so all admins get access). No code template on the pages loaded; Linear GraphQL API over HTTPS.

## Fields the listings check
- OAuth: authorize `GET https://linear.app/oauth/authorize`, token `POST https://api.linear.app/oauth/token`; callback URLs configured on the app; `scope` comma-separated: `read` (always), `write`, `issues:create`, `comments:create`, `admin` (avoid), `timeSchedule:write`; `actor=app` creates resources as the application; PKCE supported (`code_challenge`, `code_challenge_method` `plain`|`S256`). Other app form fields (name, developer URL, webhooks, public toggle) not described on the page (unconfirmed).
- Directory submission: Google Form (https://docs.google.com/forms/d/e/1FAIpQLSdYKtml_JsVv-sXuP4_eiT-l-8K6IaoEQrxdAP_7w-mb1-hSQ/viewform) with a sample page for copy style/length.
- Assets: color icon 320x320 (.svg, .jpeg or .png)

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 · recipes/notion.md (reported line 15)May include surrounding context.

md
## Fields the listings check
- Connection: name, development workspace, Redirect URI(s), installation scope **Any workspace** (only these can be listed; "cannot be changed later" is from a search summary of the docs, unconfirmed), capabilities (read / update / insert content). Client ID + secret on the Configuration tab.
- OAuth: authorize `https://api.notion.com/v1/oauth/authorize` with `owner=user`, `response_type=code`, `client_id`, `redirect_uri`; token `POST https://api.notion.com/v1/oauth/token` with HTTP Basic `CLIENT_ID:CLIENT_SECRET`, JSON `grant_type: "authorization_code"`, `code`, `redirect_uri` (when required). Response: `access_token`, `refresh_token`, `bot_id`, `workspace_id`.
- API header: `Notion-Version: 2026-03-11` on every request (unless using the JS SDK).
- Marketplace listing (Listings > Connections > Drafts): listing name and description, category and tags, listing images and logo, associated public connection. Image sizes, privacy policy/terms/support fields no

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 · recipes/notion.md (reported line 15)May include surrounding context.

md
## Fields the listings check
- Connection: name, development workspace, Redirect URI(s), installation scope **Any workspace** (only these can be listed; "cannot be changed later" is from a search summary of the docs, unconfirmed), capabilities (read / update / insert content). Client ID + secret on the Configuration tab.
- OAuth: authorize `https://api.notion.com/v1/oauth/authorize` with `owner=user`, `response_type=code`, `client_id`, `redirect_uri`; token `POST https://api.notion.com/v1/oauth/token` with HTTP Basic `CLIENT_ID:CLIENT_SECRET`, JSON `grant_type: "authorization_code"`, `code`, `redirect_uri` (when required). Response: `access_token`, `refresh_token`, `bot_id`, `workspace_id`.
- API header: `Notion-Version: 2026-03-11` on every request (unless using the JS SDK).
- Marketplace listing (Listings > Connections > Drafts): listing name and description, category and tags, listing images and logo, associated public connection. Image sizes, privacy policy/terms/support fields no