skills
- Type
- MCP server
- Scanned
- Nov 13, 2026
- Version
- 2c5958ecac7c
- Files read
- 384
This project provides guides and helper scripts that teach AI coding assistants how to create and edit Office documents and PDFs, design visuals and web apps, and build tools with the Claude API. The biggest risk is a local review web server without basic security checks that could let malicious websites or untrusted test data read your project files or tamper with results, alongside guides that can expose login tokens or automatically run code prompted by untrusted text.
What it can reach
- Files
- Reads, creates, unpacks, and overwrites project source files, skill and configuration files, evaluation logs, images, HTML pages, and Office and PDF documents in your workspace and temporary folders.
- Network
- Connects to Anthropic and cloud partner APIs, third-party tool servers (MCP), package registries, public font and script hosts, and Claude documentation sites, and runs a local web server on port 3117.
- Shell commands
- Runs command-line tools and custom scripts, including Python, Node.js, Bun, LibreOffice, PDF and document converters, compilers, Git, curl, package managers, the Claude and ant CLIs, and local development servers.
- Secrets
- Reads Anthropic and cloud provider API keys and tokens from environment variables, reads saved login profiles from ~/.config/anthropic/, and passes environment variables to child processes.
- Browser
- Opens local review pages in your default web browser and launches an automated headless Chromium browser via Playwright to test local web pages.
- System
- Stops existing processes tied to the local review server's port and installs the pnpm package manager globally if it is missing.
- Other
- Reads from and writes to connected workplace tools (such as Slack, Email, Calendar, Google Drive, SharePoint, Notion, Jira, Linear, and GitHub) when those integrations are enabled.
Findings
-
HighLocal eval review server lacks authentication, Origin, and Host checks
skills/skill-creator/eval-viewer/generate_review.py:349
When skill-creator launches its review viewer in the background on 127.0.0.1:3117, ReviewHandler accepts GET and POST requests with no password, token, Origin check, Content-Type check, or Host header check. Any website the user visits in their browser while the server is running can send a cross-origin POST request to http://127.0.0.1:3117/api/feedback to overwrite feedback.json with attacker-chosen instructions, which the agent then reads to modify the skill and run further iterations. Without Host validation, a website using DNS rebinding can also fetch GET / and steal all workspace prompts and output files embedded in the page.
Fix: Require a random per-session token on requests, validate that Host matches 127.0.0.1 or localhost on the bound port, and verify Origin and Content-Type on POST /api/feedback.
-
MediumUnescaped JSON embedded into script tag allows XSS in eval viewer
skills/skill-creator/eval-viewer/generate_review.py:70
generate_html() serializes workspace prompts, grading notes, and output files (including .html, .js, .md, and .txt files, following symlinks in outputs/) with json.dumps() and inserts the raw string directly inside a <script> tag in viewer.html. Because Python's json.dumps() does not escape closing </script> tags, any untrusted prompt or generated output containing </script> breaks out of the script block and runs arbitrary JavaScript in the user's browser when the viewer or static HTML report is opened, allowing theft of embedded files or tampering with feedback.
Fix: Escape <, >, and & as Unicode escapes (such as \u003c) in data_json before substituting into the HTML template, and refuse symlinks when scanning outputs/.
-
MediumXML External Entity (XXE) local file read in Office document validators
skills/docx/scripts/office/validators/base.py:167
Although validate.py uses defusedxml for initial checks, BaseSchemaValidator and its subclasses (duplicated across skills/docx, skills/pptx, and skills/xlsx) call lxml.etree.parse() with default parser settings. In lxml, resolve_entities=True is enabled by default and permits local file:// external entities even though network fetches are disabled. Validating an untrusted .docx or .pptx containing a crafted DOCTYPE ENTITY declaration expands local files into the XML tree and prints excerpts of those files into validation error output read by the agent, as well as enabling entity-expansion denial of service.
Fix: Pass lxml.etree.XMLParser(resolve_entities=False, no_network=True, dtd_validation=False) to every lxml.etree.parse() and fromstring() call across the docx, pptx, and xlsx validators.
-
MediumCLI flag injection when running trigger evaluation queries
skills/skill-creator/scripts/run_eval.py:72
In the claude CLI, -p (--print) is a boolean flag rather than an option taking an argument. run_eval.py passes each eval query on the command line without a -- delimiter or stdin piping. If a user runs trigger evaluation on an untrusted eval set containing a query starting with a hyphen (for example, --mcp-config=evil.json), claude interprets the query as a command-line flag and can spawn arbitrary MCP server commands defined in the config.
Fix: Send the query over stdin (as improve_description.py does) or place "--" before query in the command list, and validate skill_name so path separators cannot write outside .claude/commands/.
-
MediumFixed /tmp LibreOffice profile path allows local macro hijacking
skills/docx/scripts/accept_changes.py:16
accept_changes.py writes and executes its LibreOffice StarBasic macro inside a hardcoded /tmp/libreoffice_docx_profile directory, and _setup_libreoffice_macro() skips writing Module1.xba if the file already exists and contains the text AcceptAllTrackedChanges. On a shared multi-user machine, another local user can pre-create /tmp/libreoffice_docx_profile/user/basic/Standard/Module1.xba with malicious StarBasic commands that run under the victim user's account when accept_changes.py is invoked.
Fix: Create a fresh per-run profile directory with tempfile.TemporaryDirectory() (matching skills/xlsx/scripts/recalc.py) instead of using a fixed path under /tmp.
-
MediumFixed /tmp shared library path loaded via LD_PRELOAD
skills/docx/scripts/office/soffice.py:50
When AF_UNIX sockets are blocked, soffice.py (duplicated in skills/docx, skills/pptx, and skills/xlsx) sets LD_PRELOAD to /tmp/lo_socket_shim.so and reuses any existing file at that path without checking ownership or permissions. In any shared environment where AF_UNIX sockets are restricted, another local user can pre-create /tmp/lo_socket_shim.so to run arbitrary native code whenever LibreOffice is launched.
Fix: Compile and store lo_socket_shim.so inside a private user-owned directory (mode 0700) created with tempfile.mkdtemp() rather than directly in the shared temp directory.
-
MediumUnattended eval hillclimb loop applies and runs LLM-proposed edits derived from untrusted transcripts
skills/claude-api/shared/evals/eval-hillclimb.md:117
The hillclimb workflow feeds stored evaluation transcripts (which build-eval.md sources from untrusted production logs, customer support tickets, and live web-search tool outputs) into an analyzer subagent without framing them as untrusted data, asks for a diff to the target artifact (which Step 1 allows to include harness or agent loop code and helper scripts), defaults approve_each_round to false, and asks the user to session-allowlist the eval runner command so rounds execute unattended. Because the runner hash check only covers paths explicitly listed in _state.json.harness_paths (not editable app code invoked by the runner), prompt injection in a transcript can steer the agent into writing malicious code and immediately executing it without per-round user review.
Fix: Default approve_each_round to true whenever executable code, helper scripts, or tool definitions are in scope, and explicitly instruct the analyzer subagent to treat all transcript and dataset fields as untrusted data rather than instructions.
-
LowPath traversal in Python MCP resource handler reference example
skills/mcp-builder/reference/python_mcp_server.md:539
The Python MCP implementation guide teaches the coding agent to expose file://documents/{name} resources by interpolating {name} directly into ./docs/{name} and opening the file without resolving the path or verifying it stays inside ./docs. MCP servers generated from this template allow callers to read arbitrary files outside the intended directory using .. path segments.
Fix: Update the reference example to resolve the candidate path with Path.resolve() and verify it is relative to the resolved ./docs root before reading.
-
LowUnsanitized filenames in Managed Agents file download examples
skills/claude-api/python/managed-agents/README.md:343
The Managed Agents Python and TypeScript README examples save session output files directly to f.filename in the host working directory without stripping directory components via os.path.basename() or writing to a dedicated output folder (contradicting the path-traversal warning in shared/tool-use-concepts.md). Code generated from this pattern lets a prompt-injected session overwrite sensitive files on the host if an output filename contains traversal segments or dotfiles.
Fix: Sanitize f.filename with os.path.basename() (and path.basename() in TypeScript) and join it under a dedicated output directory in the example snippets.
-
LowGlobal and unpinned npm package installs at runtime
skills/web-artifacts-builder/scripts/init-artifact.sh:36
init-artifact.sh automatically installs pnpm globally (npm install -g pnpm) if missing and, together with bundle-artifact.sh, downloads and installs dozens of unpinned npm packages (such as create-vite, parcel, html-inline, and @radix-ui/*) at runtime with lifecycle scripts enabled. Modifying the global Node environment is not mentioned in SKILL.md, and unpinned installs expose builds to compromised upstream releases.
Fix: Use npx/corepack instead of global npm install -g, pin dependency versions (or ship a lockfile), and pass --ignore-scripts where native postinstall hooks are unnecessary.
-
LowCustomer-facing Managed Agent quickstart templates enable unrestricted shell and web tools alongside enterprise MCPs
skills/claude-api/shared/managed-agents-quickstarts/support-to-eng-escalator.md:46
The support-to-eng-escalator quickstart (along with incident-commander.md, support-agent.md, and contract-tracker.md) ingests untrusted external content such as customer Intercom messages, logs, and Sentry payloads, and instructs the agent to run customer-supplied reproduction steps inside the container. Unlike bug-hunter.md and watchlist-scanner.md, these templates enable the full agent_toolset_20260401 (including bash, web_search, and web_fetch) without setting default_config to disabled, omit an environment.yaml to restrict outbound network access, omit untrusted-data prompt boundaries, and attach authenticated MCP servers for Atlassian/Jira, Slack, Intercom, GitHub, Linear, Notion, and Box. Prompt injection in a ticket or alert can cause the cloud agent to read internal workspace data via MCP and exfiltrate it over the network.
Fix: Set default_config: {enabled: false} on agent_toolset_20260401 to enable only required tools, add an environment.yaml with limited networking and explicit allowed_hosts, configure approval policies on external or write MCP actions, and instruct the agent that external tickets and payloads are data, never instructions.
-
LowHigh $500 USD per-session default spending limit in Managed Agent quickstart configs
skills/claude-api/shared/managed-agents-quickstarts/field-monitor.md:65
Six Managed Agent quickstart templates (field-monitor.md, sprint-retro-facilitator.md, bug-hunter.md, explainer-video-maker.md, performance-tuner.md, and watchlist-scanner.md) configure max_list_cost at $500 USD per session—including scheduled cron deployments that fire automatically every week or sprint—even though normal runs are estimated at only a few dollars. If deployed unchanged, a looping multi-agent run or prompt injection on scraped external pages can consume up to $500 in API spend in a single firing.
Fix: Lower the default max_list_cost in quickstart session and scheduled deployment YAMLs to a small cap matching normal run cost (such as $10 to $25) and ask the user before setting higher limits.
-
LowInternal communications guides automatically scan Slack, Email, Drive, and Calendar without confirmation
skills/internal-comms/examples/3p-updates.md:15
Although internal-comms/SKILL.md describes itself as company formatting guidelines, its sub-guides (3p-updates.md, company-newsletter.md, and faq-answers.md) tell the agent to proactively search wide-reaching Slack channels, Email, Google Drive, Calendar, and external press without asking confirmation first. Beyond reading sensitive messages the user may not expect a formatting skill to touch, ingesting broad channels, emails, and press articles exposes the agent to prompt injection from untrusted messages.
Fix: Require explicit user confirmation and scope selection (such as specific channels or documents) before searching connected tools, matching the safeguard in skills/doc-coauthoring/SKILL.md.
-
LowBundled Office XSD schemas import external definitions over cleartext HTTP
skills/docx/scripts/office/schemas/ecma/fouth-edition/opc-coreProperties.xsd:8
Unlike the other bundled Office schemas that were edited to use local relative paths, opc-coreProperties.xsd (and identical copies under skills/pptx/ and skills/xlsx/) imports Dublin Core schemas from http://dublincore.org over unencrypted HTTP with no local copy present. If an XML schema validator resolves these imports over the network when validating a document's core.xml metadata, an on-path network attacker could intercept and modify the imported schema.
Fix: Bundle dc.xsd and dcterms.xsd locally alongside opc-coreProperties.xsd in docx, pptx, and xlsx, and update schemaLocation to relative local paths.
