skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- d2ef2680594f
- Files read
- 620
This project is a collection of skills and helper scripts that teach AI agents how to handle coding, testing, media production, document drafting, and research tasks. The biggest risk is that several skills and background scripts bypass normal safety prompts to run commands, edit files, and install unverified packages automatically, which could let untrusted content change files or run unwanted commands on your computer.
What it can reach
- Files
- Reads and writes project files, documents, images, videos, and hidden configuration or state folders in your home directory (such as ~/.claude/, ~/.ecc/, and ~/.local/).
- Network
- Connects to software package registries (npm, PyPI, GitHub), user-provided web links, and third-party services including OpenAI, Gemini, X, Jira, Stripe, ElevenLabs, fal.ai, and VideoDB.
- Shell commands
- Runs terminal commands, Python and Node scripts, package installers (like npm, pip, and Homebrew), developer and media tools (like git, Docker, ffmpeg, Pandoc, and Blender), and background AI agent processes.
- Secrets
- Reads API keys, login tokens, SSH credentials, and payment wallet keys from environment variables, .env files, and your computer's keychain.
- Browser
- Controls automated Chrome or Chromium browsers to inspect and scrape web pages, test user interfaces, prepare e-signature drafts, export slide PDFs, and record videos.
- System
- Sets up scheduled background tasks, sends desktop notifications, records macOS screen and audio, automates Windows desktop clicks and typing, opens terminal windows, and uses local mail or messaging apps.
Findings
-
HighUnsandboxed host shell and file permissions in skill-comply scenario runner
skills/skill-comply/scripts/runner.py:98
When evaluating a skill file, skill-comply generates test prompts from that skill's text and runs a non-interactive claude -p subprocess with pre-approved Bash, Write, and Edit tools on the host machine. Although setup commands are confined to a temporary folder, the Claude run itself is not sandboxed, so a prompt-injected skill file can run arbitrary shell commands without user confirmation.
Fix: Run scenario execution inside an isolated container or VM, or remove Bash from --allowedTools and restrict file tools strictly to sandbox_dir.
-
MediumVisa translation skill disables user confirmation while installing unpinned packages and running generated Python scripts on OCR text
skills/visa-doc-translate/SKILL.md:10
This skill instructs the agent to run every step automatically without asking the user for confirmation, including installing unpinned packages (easyocr, pytesseract, pillow, reportlab, tesseract) and writing and executing a new Python script containing text extracted via OCR from an input image. Because OCR text from third-party certificates or images is not treated as untrusted data, prompt-injection text or unescaped quotes inside a document image can trick the agent into generating and running arbitrary Python code on the user's computer.
Fix: Require explicit user confirmation before installing dependencies or running scripts, pin package versions, and pass OCR/translated text as pure data (such as JSON) to a fixed PDF script instead of generating and executing Python code.
-
MediumRepository-local agent files shadow trusted global and plugin agents in team-builder
skills/team-builder/SKILL.md:58
When discovering agent personas, team-builder loads ./agents/**/*.md and ./agents/*.md from the current project directory before ~/.claude/agents/ and gives local user files precedence over plugin agents when names collide (for example, ./agents/security-reviewer.md overrides ecc:security-reviewer). If a user runs team-builder inside a cloned or untrusted repository, a malicious markdown file in ./agents/ can impersonate a trusted reviewer and inject its raw text directly as the prompt for a general-purpose subagent with full tool access.
Fix: Give ~/.claude/agents/ and verified plugin agents precedence over repository-local ./agents/ files, label repository-local personas clearly, and require user confirmation before running them.
-
MediumUnrestricted file Read and Write permissions in background observer daemon
skills/continuous-learning-v2/agents/observer-loop.sh:400
When the continuous-learning-v2 observer is enabled, observe.sh logs tool outputs (including untrusted web pages, issues, or repo files) and observer-loop.sh feeds them to an unattended background Claude Haiku process with --allowedTools Read,Write. Because those tools are not path-scoped to the instincts folder, prompt injection in logged content can trick the model into reading sensitive files or overwriting user startup and configuration files.
Fix: Restrict --allowedTools to Read on the analysis file and Write inside INSTINCTS_DIR, or return structured JSON and write instinct files in deterministic host code.
-
MediumUnescaped counterparty fields and unsandboxed Pandoc DOCX build
skills/master-agreement-generator/scripts/build-agreement.js:182
While schedule table cells are entity-encoded, counterparty fields such as legal, addr, juris, signer, and supplement are inserted raw into Markdown and passed to pandoc without --sandbox. Markdown images or raw openxml blocks in counterparty-supplied fields can inject hidden contract markup, trigger outbound HTTP requests, or embed local files into the generated DOCX draft.
Fix: Escape Markdown/HTML syntax in all spec string fields and pass --sandbox to pandoc during DOCX conversion.
-
MediumHTTP redirect and DNS rebinding bypass in remote instinct import
skills/continuous-learning-v2/scripts/instinct-cli.py:225
_validate_import_url enforces HTTPS and blocks private or loopback IP addresses, but urllib.request.urlopen follows HTTP 3xx redirects by default without re-validating the redirect target. A public HTTPS URL can redirect to plaintext HTTP, localhost, or cloud metadata endpoints, bypassing the SSRF and transport checks.
Fix: Use a custom HTTPRedirectHandler that re-runs _validate_import_url on every redirect hop (or block redirects) and verify the connected peer IP.
-
LowDjango pre-deployment verification checklist creates a superuser with password admin
skills/django-verification/SKILL.md:155
This skill is designed for pre-PR and pre-deployment verification on staging or production environments where DEBUG is False and a real database is configured. Phase 6 includes a command that creates a Django superuser (admin@example.com) with the hardcoded weak password 'admin'. Running Phase 6 against a staging, shared, or production database plants a live administrative account with a trivial password.
Fix: Remove superuser creation from the pre-deployment verification commands and create test admin accounts only inside isolated, auto-rolled-back test fixtures.
-
LowSetup and scan skills download and execute unpinned packages at runtime
skills/configure-ecc/SKILL.md:189
When CLAUDE_PLUGIN_ROOT is unavailable or when configuring Kimi, configure-ecc instructs the agent to run npx --yes --package ecc-universal without a pinned version, even during the pre-confirmation dry-run preview step. Because --yes bypasses the install prompt, npm downloads and executes whatever code is currently published as latest. Similar unpinned runtime downloads appear in skills/security-scan/SKILL.md (npx ecc-agentshield scan .), skills/angular-developer/SKILL.md (npx @angular/cli@latest), skills/unified-memory/SKILL.md (npm install -g ecc-universal), and skills/ito-compute/SKILL.md (unpinned git clone followed by npm ci).
Fix: Pin exact reviewed package versions and immutable git commit SHAs, and ask for user confirmation before downloading or executing remote packages.
-
LowMCP configuration snippets use unpinned npx -y packages alongside secret API keys
skills/codehealth-mcp/SKILL.md:48
The MCP setup snippet runs npx -y @codescene/codehealth-mcp without pinning a version while passing CS_ACCESS_TOKEN in env and granting access to local repository files. The same unpinned npx -y MCP pattern is used in skills/exa-search/SKILL.md (exa-mcp-server with EXA_API_KEY), skills/fal-ai-media/SKILL.md (fal-ai-mcp-server with FAL_KEY), and skills/nutrient-document-processing/SKILL.md (@nutrient-sdk/dws-mcp-server with NUTRIENT_DWS_API_KEY and SANDBOX_PATH). Future sessions can silently pull a newly published package version with immediate access to those credentials.
Fix: Pin exact reviewed package versions (such as @codescene/codehealth-mcp@x.y.z) in every MCP server configuration example.
-
LowDouble-quoted shell placeholders allow command substitution from project descriptions
skills/opensource-pipeline/SKILL.md:244
Step 1 asks the agent to analyze the source project to suggest a README description, and Step 7 interpolates {description} inside double quotes in a Bash command. In Bash, double quotes still expand backticks and $() command substitutions, so a description containing Markdown code backticks or crafted metadata from an analyzed project can execute unintended shell commands when publishing. Similarly, skills/ck/SKILL.md leaves [arg] and [name] unquoted in Bash command templates.
Fix: Single-quote literal values (escaping apostrophes) or pass descriptions through a file or quoted heredoc rather than double-quoted shell strings.
-
LowFlox host environment template binds auto-reloading Uvicorn server to 0.0.0.0
skills/flox-environments/SKILL.md:399
Flox environments run directly on the developer's host OS rather than inside an isolated container. The Full-Stack Example defines a serve() shell function that starts uvicorn with --reload bound to 0.0.0.0:8000 instead of 127.0.0.1, exposing the unauthenticated development server to other devices on the local network whenever a user or agent copies this template and runs serve.
Fix: Bind local development servers on the host OS to 127.0.0.1 by default instead of 0.0.0.0.
-
LowUnpinned runtime npm install on every slide PDF export
skills/frontend-slides/scripts/export-pdf.sh:375
Every run of export-pdf.sh creates a fresh temporary directory and executes an unpinned npm install playwright without a lockfile or --ignore-scripts, downloading and executing third-party package code at runtime.
Fix: Pin an exact playwright version with a lockfile or require Playwright to be pre-installed rather than installing unpinned packages in a temp folder at runtime.
-
LowBash arithmetic command injection and unvalidated paths in iconify_gen.sh
skills/ios-icon-gen/scripts/iconify_gen.sh:164
Unlike the sibling Swift script, iconify_gen.sh does not validate --size before evaluating it inside Bash arithmetic $((SIZE * 2)), where array-subscript expressions such as a[$(cmd)] execute shell commands, nor does it block path traversal in icon_id or asset_name.
Fix: Validate that SIZE is a positive integer, COLOR is a 6-digit hex string, and icon_id and asset_name contain no path separators or '..' segments.
