← MCP server and skill grades

skills

github/awesome-copilot/skills

F
Type
Skill
Scanned
Nov 13, 2026
Version
222eba62d518
Files read
1546

This collection of skills and helper scripts gives AI coding assistants step-by-step workflows for software development, cloud services, testing, document conversion, and codebase audits. The biggest risk is that some skills download and run unverified code from the internet without asking or execute malicious code hidden inside untrusted project files and datasets, while others expose local files on passwordless servers or turn off safety prompts.

What it can reach

Files
Reads, creates, modifies, and deletes project files, Office and PDF documents, diagrams, HTML reports, Desktop files, shell startup profiles, parent-folder Git worktrees, and user-level AI config folders (such as ~/.copilot, ~/.claude, and ~/.codex).
Shell commands
Runs terminal commands, build and test tools, Git, Docker, cloud CLIs (AWS, Azure, GitHub), package installers (npx, pip, uv), and bundled Python, Node.js, Bash, and PowerShell scripts.
Network
Downloads packages from GitHub, npm, PyPI, and NuGet, connects to cloud, AI, database, and third-party services (including GitHub, Azure, AWS, OpenAI, Gemini, OpenRouter, Arize, and Namecheap), and runs local web servers.
Secrets
Reads API keys, database passwords, and login tokens from environment variables, local config files (such as ~/.namecheap-api), cloud and GitHub CLI sessions, and GitHub browser cookies.
Browser
Launches and controls local Chrome, Edge, Firefox, and WebKit browsers (including saved login profiles for Microsoft Partner Center), exports diagrams, tests web pages, and opens local HTML reports.
System
Reads and writes the clipboard, captures desktop screenshots, inspects running processes and Windows developer tools, installs system browser dependencies, and controls desktop apps like Adobe Illustrator and FreeCAD.

Findings

1 critical9 high17 medium25 low
  1. CriticalAutomatic fallback to pickle deserialization runs arbitrary code on dataset load

    skills/datanalysis-credit-risk/references/func.py:40

    get_dataset() ignores the file extension and loops through read_parquet, read_csv, read_excel, and finally pd.read_pickle() inside a bare try/except block. Because binary pickle payloads fail Parquet, UTF-8 CSV, and Excel parsing, any malicious dataset file—even one named .csv or .xlsx—falls through to pd.read_pickle() and executes arbitrary code on the user's machine.

    Fix: Remove pd.read_pickle from get_dataset(), dispatch strictly by validated file extension for safe formats (.parquet, .csv, .xlsx), and never unpickle untrusted data.

  2. HighScript injection via unescaped CSV comments in auto-opened HTML report

    skills/geofeed-tuner/SKILL.md:684

    Phase 2 accepts geofeed CSVs from untrusted files or remote URLs, and Phase 3 saves comment lines verbatim into comments.json. Phase 5 instructs the agent to insert json.dumps(comments) directly into a HTML script block and states that no extra escaping is needed, then automatically opens the HTML file in the user's default browser. Because Python's json.dumps does not escape closing script tags, a CSV comment containing </script><script>... executes arbitrary JavaScript in the user's browser under the file:// origin.

    Fix: Escape angle brackets and ampersands as Unicode escape sequences after json.dumps(comments) before substituting into the HTML script block.

  3. HighCopilot SDK templates auto-approve all tool and shell permission requests

    skills/copilot-sdk/SKILL.md:66

    Every createSession code sample across TypeScript, Python, Go, and .NET configures onPermissionRequest: approveAll (or PermissionHandler.approve_all) without restricting availableTools or warning what approveAll permits. Because the Copilot SDK runtime exposes shell commands, file edits, and MCP tools by default, applications built from these examples allow prompt injection from untrusted user input, files, or MCP responses to execute arbitrary commands on the host.

    Fix: Replace approveAll in templates with a deny-by-default or explicit allowlist permission handler paired with availableTools, and document the security risk of approveAll.

  4. HighBackground HTTP server exposes working directory without password or Host check

    skills/daily-focus-board/scripts/serve-board.ps1:15

    serve-board.ps1 starts Python's built-in http.server hidden in the background on 127.0.0.1:8799 serving the entire working directory ($Dir defaults to the current directory) with directory listings enabled and no authentication or Host/Origin validation. While running, any website the user visits in their browser can use DNS rebinding to list and read arbitrary files in that directory (such as .env files, source code, or personal board data).

    Fix: Serve only the single generated HTML file from an isolated directory using a server that validates the Host header, disables directory listings, and requires a random per-session token.

  5. HighUnescaped PR content enables script injection in local dashboard HTML

    skills/pr-dashboard/scripts/pr-dashboard-cli.mjs:235

    pr-dashboard-cli.mjs fetches GitHub pull requests (including PRs from strangers requesting review or mentioning the user), embeds JSON.stringify(md) containing raw PR titles and descriptions directly inside an inline <script> block in /tmp/pr-dashboard.html, and opens it in the user's default browser. Because JSON.stringify does not escape </script> (and String.prototype.replace expands $ patterns), a PR containing </script><script>... executes arbitrary JavaScript in the file:// page and can exfiltrate private PR data shown on the dashboard.

    Fix: Escape '<' as '\u003c' in the serialized JSON, pass a function (() => ...) as the second argument to .replace(), and write the output file inside a private fs.mkdtempSync directory.

  6. HighMSBuild property injection via untrusted Directory.Packages.props or packages.config

    skills/winmd-api-search/scripts/Update-WinMdCache.ps1:138

    Although SKILL.md states cache generation is safe on a fresh clone without building the target repo, Update-WinMdCache.ps1 reads the Microsoft.WindowsAppSDK version attribute from the target repo's Directory.Packages.props or packages.config without validation and passes it in -p:WinAppSdkVersion=$winAppSdkVersion to dotnet restore and dotnet build. Because MSBuild splits -p: arguments on semicolons, a crafted version string can inject properties like CustomAfterMicrosoftCommonTargets to execute arbitrary commands from the cloned repo.

    Fix: Validate $winAppSdkVersion against a strict NuGet version regex (such as ^[0-9]+\.[0-9]+\.[0-9]+(-[0-9A-Za-z.-]+)?$) before passing it to dotnet commands.

  7. HighArbitrary local file inclusion via image path traversal in Markdown-to-DOCX converter

    skills/md-to-docx/scripts/md-to-docx.mjs:331

    When converting Markdown to Word, md-to-docx.mjs resolves image links with resolve(inputDir, href) without confining them to inputDir or verifying image magic bytes (non-PNG files get fallback dimensions and are still embedded raw into the .docx zip). Converting an untrusted Markdown file with links like ![](../../.env) or ![](/Users/me/.ssh/id_rsa) silently packs those local files into the generated .docx archive.

    Fix: Resolve symlinks with realpathSync, verify the target path stays inside inputDir, and check that the file has an allowed image extension and valid image header bytes.

  8. HighDestructive shutil.rmtree on output directory in Word and PDF converters

    skills/convert-word-to-md/scripts/convert_word_to_md.py:206

    Both convert_word_to_md.py and convert-pdf-to-md/scripts/convert_pdf_to_md.py delete the entire destination directory with shutil.rmtree(dest_dir) if it already exists. Passing an existing folder to -o (such as -o . or -o ./docs), or batch-converting a folder containing a file like src.docx alongside an existing src/ directory, permanently wipes out the existing directory tree.

    Fix: Do not call shutil.rmtree(dest_dir); refuse to overwrite non-empty target directories or only remove the generated img/ subdirectory as done in convert_excel_to_md.py.

  9. HighScript injection in unsandboxed headless browser and unpinned remote script load

    skills/drawio/scripts/drawio-to-png.mjs:44

    When the draw.io desktop CLI is absent, drawio-to-png.mjs launches Chrome/Edge with --no-sandbox, interpolates raw .drawio file contents into an inline <script> block without escaping </script>, and loads an unpinned remote script from viewer.diagrams.net without Subresource Integrity. A .drawio file containing </script><script>... executes arbitrary JavaScript inside an unsandboxed local browser instance.

    Fix: Pass diagram XML via page.evaluate arguments rather than HTML string interpolation, vendor or pin viewer-static.min.js with an integrity hash, block unexpected network requests, and drop --no-sandbox where supported.

  10. HighAudit workflow executes scripts and instructions from the target repository root first

    skills/quality-playbook/phase_prompts/phase6.md:8

    When auditing a codebase, quality-playbook resolves SKILL.md and quality_gate.py by checking the current working directory (the repository being audited) before checking the user's installed skill directory, and also runs quality/mechanical/verify.sh and python3 -m bin.quality_playbook from the target directory. Auditing an untrusted repository that contains a malicious quality_gate.py, verify.sh, bin/quality_playbook.py, or root SKILL.md will execute the attacker's code or instructions as trusted playbook infrastructure.

    Fix: Resolve SKILL.md, quality_gate.py, and helper modules exclusively from the trusted skill installation path rather than the target repository root, and never execute pre-existing scripts inside quality/mechanical/ without explicit user review.

  11. MediumAutomatic clone and execution of unpinned third-party installer on generic prompts

    skills/ponder/SKILL.md:14

    This skill triggers on broad everyday phrases such as 'think this over properly' or 'audit/rate/grade every X' without mentioning external software in its description. If the ponder binary is not installed, it instructs the agent to clone https://github.com/Quinnod345/ponder at unpinned HEAD and run install.sh without asking for user approval or verifying a commit hash. Any compromise of that personal repository executes arbitrary shell code on the user's machine.

    Fix: Require explicit user consent before cloning or running install.sh, pin the repository to a reviewed commit SHA, and disclose the external binary requirement in the frontmatter.

  12. MediumSilent runtime self-update of skill and Python package via setup.sh

    skills/eval-driven-dev/SKILL.md:34

    Before starting and on every restart, the skill instructs the agent to run setup.sh, which updates both the eval-driven-dev skill itself and the pixie-qa Python package to their latest remote versions and starts a background web server. It also tells the agent to install missing Python packages on its own without asking the user. Self-updating at runtime bypasses whatever version of the skill and package the user reviewed at install time.

    Fix: Remove automatic self-updates from setup.sh, pin pixie-qa to an exact reviewed version, and ask user confirmation before installing packages or starting background servers.

  13. MediumUnpinned third-party CLI extension auto-installed with access to GitHub session cookies

    skills/gh-attach/SKILL.md:13

    The Prerequisites step automatically installs sudosubin/gh-attach from the latest unpinned commit on GitHub without asking the user first, and then runs it to upload files using the user's GitHub browser user_session cookie. Because a GitHub browser session cookie acts as a full account credential, a compromise of the unpinned upstream repository would expose full GitHub account access.

    Fix: Pin the extension to a reviewed release tag or commit SHA with --pin and ask for explicit user confirmation before installing it or accessing browser session cookies.

  14. MediumGenerated Ktor SSE MCP server binds to 0.0.0.0 without authentication

    skills/kotlin-mcp-server-generator/SKILL.md:426

    The SSE transport template starts Ktor's embeddedServer(Netty, port = 8080) without setting host = '127.0.0.1', which defaults in Ktor to listening on 0.0.0.0 across all network interfaces, and includes no authentication or Origin header validation. Any server generated from this snippet exposes its MCP tools to anyone on the local network and to cross-origin browser requests.

    Fix: Bind embeddedServer explicitly to host = "127.0.0.1" by default, validate Origin headers, and require authentication before exposing SSE or HTTP transports.

  15. MediumUndisclosed MCP calls write workspace convention files and report telemetry

    skills/integrate-context-matic/SKILL.md:84

    This skill triggers whenever a user asks to integrate any third-party API or SDK, and its description lists only fetch_api, ask, model_search, and endpoint_search. However, the instructions also direct the agent to call add_guidelines and add_skills (writing persistent skill and workflow files into the repository) and update_activity (sending developer milestone telemetry to the ContextMatic MCP server) without asking user consent.

    Fix: Disclose add_guidelines, add_skills, and update_activity in the skill description, narrow the trigger to explicit ContextMatic requests, and ask permission before writing files or sending telemetry.

  16. MediumUndisclosed transmission of geofeed location entries to external MCP server

    skills/geofeed-tuner/SKILL.md:594

    Although the frontmatter presents this skill as a local Python CSV validator with no mention of external network services, the mandatory Phase 4 step sends all deduplicated country, region, and city entries from the user's geofeed CSV to https://mcp.fastah.ai/mcp without asking permission. Operators tuning unpublished geofeeds may inadvertently leak planned PoP or infrastructure locations to a third party.

    Fix: Document the external https://mcp.fastah.ai/mcp call in the skill description and compatibility metadata, and ask user consent before sending geofeed entries over the network.

  17. MediumUnpinned npx execution from mutable GitHub branch in readiness skills

    skills/acreadiness-assess/SKILL.md:25

    acreadiness-assess (as well as acreadiness-generate-instructions and acreadiness-policy) executes npx -y github:microsoft/agentrc directly from the mutable default branch of the repository without version or commit pinning and with confirmation suppressed (-y). If the upstream repository branch is compromised, running an assessment executes untrusted code locally.

    Fix: Pin the package reference to an immutable commit SHA or published npm version and remove -y so first-time package execution requires confirmation.

  18. MediumRuntime execution of instructions fetched from external mutable URLs

    skills/arize-instrumentation/SKILL.md:108

    The skill directs the agent to fetch https://arize.com/docs/PROMPT.md and linked integration pages at runtime and follow their package-installation and code-modification instructions, and also recommends running arize-tracing-assistant@latest. Because remote web content can change or be tampered with via CDN or DNS compromise, treating live web pages as executable agent instructions introduces a remote prompt-injection and supply-chain risk.

    Fix: Bundle reviewed installation and instrumentation guidance locally in the skill's references directory, treat fetched web content as untrusted reference data only, and pin MCP package versions.

  19. MediumSkill instructs agent to execute shell commands found inside code comments

    skills/quasi-coder/SKILL.md:309

    quasi-coder tells the agent to parse '()=>' shorthand comments inside source files, execute non-code actions requested there ('run commands, create files, fetch data'), and delete the shorthand lines afterward. Processing a file from an untrusted PR or collaborator lets embedded comments trigger arbitrary command execution and erase themselves from the file.

    Fix: Treat shorthand comments strictly as code-generation specifications and forbid running shell commands or network requests from in-file shorthand without explicit user approval.

  20. MediumUnescaped SQL identifiers allow SQL injection in table reconciliation

    skills/sql-server-table-reconciliation/scripts/reconcile.py:155

    Although SKILL.md states user input is never f-string interpolated into SQL, reconcile.py interpolates table names, primary-key columns, and comparison columns directly into SELECT queries without bracket-quoting or identifier validation. Crafted --tables, --primary-key, or --columns arguments (or malicious table/column names in an untrusted database) can inject arbitrary SQL statements.

    Fix: Validate schema, table, and column identifiers against INFORMATION_SCHEMA metadata and bracket-quote every identifier (escaping ']' as ']]') before building SQL queries.

  21. MediumUnescaped slide title and font name cause stored XSS on GitHub Pages

    skills/publish-to-pages/scripts/convert-pptx.py:286

    convert-pptx.py copies the first text shape on slide 1 directly into <title>{title}</title> without HTML escaping (and interpolates font names unescaped into style attributes), then publish.sh deploys the HTML publicly by default to USERNAME.github.io. Converting a third-party PPTX containing </title><script>... creates stored XSS on the user's shared github.io origin.

    Fix: Escape title, font names, and table cell text with html.escape() before inserting them into the generated HTML.

  22. MediumTLS certificate verification unconditionally disabled for SQL Server connections

    skills/sql-server-table-reconciliation/scripts/reconcile.py:46

    connect() hardcodes TrustServerCertificate=yes for both SQL-authenticated and Microsoft Entra connections, including production servers. An attacker on the network path can intercept the connection with a self-signed certificate to capture credentials or read and tamper with table data.

    Fix: Set TrustServerCertificate=no by default and only allow trusting self-signed certificates when an explicit opt-in flag is provided for local development.

  23. MediumSkill instructs proactive background exposure of localhost apps to external cloud

    skills/scoutqa-test/SKILL.md:26

    Both the frontmatter description and instructions tell the agent to proactively run scoutqa in the background after implementing web features, including against localhost URLs that ScoutQA tunnels to its remote cloud infrastructure. This sends local application pages and data outside the machine without the user asking for external testing.

    Fix: Only run ScoutQA when explicitly requested by the user, require confirmation before exposing localhost servers to remote infrastructure, and pin @scoutqa/cli to a specific version.

  24. MediumPersistent unauthenticated remote debugging port on Partner Center browser profile

    skills/microsoft-store-publisher/scripts/partner-center-control.mjs:160

    The launch command opens Edge with the user's authenticated Partner Center profile and binds --remote-debugging-port=9223 on 127.0.0.1 indefinitely. Because Chrome DevTools Protocol has no authentication (and the script's goto command is also exempt from PARTNER_CENTER_ALLOW_WRITE), any local process can take control of the signed-in Partner Center session or read its cookies.

    Fix: Avoid persistent TCP debugging ports by using pipe-based automation (--remote-debugging-pipe) within a single session, and restrict goto to approved Partner Center domains.

  25. MediumSetup script silently self-updates global and project skills via unpinned npx

    skills/eval-driven-dev/resources/setup.sh:15

    setup.sh runs an unpinned 'npx skills update ... -g -y' and '-p -y' before installing dependencies, downloading and executing the latest npm package and overwriting both global and project skill files from the mutable remote repository HEAD without prompting.

    Fix: Remove automatic global/project self-updates from setup.sh (or require explicit user opt-in) and pin CLI package versions.

  26. MediumPath traversal in WinMD cache generator when indexing untrusted projects

    skills/winmd-api-search/scripts/cache-generator/Program.cs:157

    When CacheGenerator scans a project directory, package IDs and versions read from obj/project.assets.json or packages.config are passed directly into Path.Combine without checking for '..' segments or rooted paths. In addition, namespace strings read from .winmd metadata only have '.' replaced with '_' (leaving '/' and '\' intact), so a crafted project or .winmd file in an untrusted repository can create directories and overwrite .json files outside the intended output directory.

    Fix: Reject package IDs, versions, and namespace names containing path separators or drive prefixes, and verify that Path.GetFullPath of every output path stays within Path.GetFullPath(outputDir) + Path.DirectorySeparatorChar.

  27. MediumUnescaped repository data in AI-readiness HTML report template enables XSS

    skills/acreadiness-assess/report-template.html:219

    The readiness assessment workflow scans a repository (including its agentrc.config.json and file names) and instructs the reporter agent to substitute {{rawJsonCompact}} directly inside a <script type="application/json"> block and {{rawJsonPretty}} inside a <pre> block in reports/index.html without escaping rules. If an audited repository contains a string with '</script><script>...' or raw HTML tags in its config or file names, opening reports/index.html in a browser executes arbitrary JavaScript.

    Fix: HTML-escape all text placeholders and escape '<' as '\u003c' inside {{rawJsonCompact}} before writing reports/index.html, and include a strict Content-Security-Policy meta tag.

  28. LowGitHub Actions script injection in generated issue-creation workflow

    skills/breakdown-plan/SKILL.md:421

    The automated issue-creation workflow template interpolates github.event.inputs.feature_name and github.event.inputs.epic_issue directly into an actions/github-script JavaScript literal instead of passing them via environment variables, and omits a permissions block. If adopted into a repository, crafted workflow_dispatch inputs containing backticks or template expressions execute arbitrary JavaScript in the runner with the workflow's GITHUB_TOKEN.

    Fix: Pass workflow inputs via step env variables and read process.env inside the script, and declare minimal permissions (issues: write).

  29. LowOnboarding collects raw API keys in chat and writes them to shell profiles

    skills/generate-image/SKILL.md:37

    When no API key is found, the onboarding workflow asks the user to paste their OpenAI or Gemini API key directly into the chat—exposing the secret to the model provider and chat logs—and tells the agent to persist the plaintext key into the user's shell startup profile without asking permission to edit shell configuration files.

    Fix: Instruct users to set SKILL_IMAGE_GEN_OPENAI_KEY or SKILL_IMAGE_GEN_GEMINI_KEY in their own terminal or secret manager rather than pasting keys into chat or modifying shell profiles automatically.

  30. LowNoSQL query injection in Cosmos DB stored procedure template

    skills/cosmosdb-datamodeling/SKILL.md:882

    The unique-email stored procedure example interpolates userData.email directly into a Cosmos DB SQL query string instead of using parameterized queries. Code generated from this pattern allows untrusted email input to alter the query predicate and bypass uniqueness checks or read unintended partition data.

    Fix: Replace string interpolation with a parameterized query object passing @email in the parameters array of container.queryDocuments.

  31. LowHardcoded weak database passwords and all-interface Docker port mappings

    skills/create-spring-boot-kotlin-project/SKILL.md:102

    This scaffolding skill instructs the agent to hardcode the password 'rootroot' for Redis, PostgreSQL, and MongoDB in application.properties and docker-compose.yaml while publishing ports 6379, 5432, and 27017 on all host interfaces (0.0.0.0 by default in Docker). Running the generated Compose stack exposes three databases with a trivial password to the local network.

    Fix: Bind Docker ports to 127.0.0.1 and read randomly generated database credentials from untracked environment variables.

  32. LowUnpinned remote install scripts piped directly into shells

    skills/aspire/SKILL.md:102

    skills/aspire/SKILL.md (along with azure-container-registry-cli, azure-devops-cli, and nano-banana-pro-openrouter) instructs piping unpinned remote scripts from URLs directly into bash, sudo bash, or PowerShell iex without checksum or signature verification.

    Fix: Use OS package managers or download versioned release archives and verify SHA-256 checksums before execution.

  33. LowBeginner tutorial instructs users to prompt 'Delete all files in this directory'

    skills/copilot-cli-quickstart/SKILL.md:278

    Lesson S3 tells beginners to test permission prompts by asking Copilot to delete all files in the current directory, right after Lesson S2 introduced 'Allow for session' for file operations. If the user selected 'Allow for session' earlier or enabled auto-approval, the command executes without prompting and deletes their files.

    Fix: Use a harmless test prompt (such as creating a dummy temporary file and denying it) instead of a destructive bulk-delete command.

  34. LowAPI keys passed directly on the command line

    skills/arize-ai-provider-integration/SKILL.md:113

    Provider integration commands pass secret API keys as CLI flags (--api-key $OPENAI_API_KEY; similar secret flags appear in azure-devops-cli and msstore-cli), which exposes expanded secrets in the local process table and command logs.

    Fix: Pass secrets via standard input or environment variables read directly by the CLI rather than command-line flags.

  35. LowHTML/script injection in generated Azure diagrams and unescaped path in node -e

    skills/azure-architecture-autopilot/scripts/generator.py:743

    generator.py embeds unescaped title, JSON blobs (without escaping </script>), and node names/details (via innerHTML) into the output HTML, which cli.py also opens under file:// in Puppeteer for PNG export while interpolating unescaped file paths into a single-quoted node -e script. Malicious resource names or file paths in IaC inputs can execute arbitrary JavaScript during rendering or viewing.

    Fix: HTML-escape title and tooltip/sidebar fields (or use textContent), escape '</' in embedded JSON, and pass file paths to Node via process.argv instead of string interpolation in node -e.

  36. LowPlaintext database passwords written to workspace XML and echoed to chat

    skills/ssma-console/SKILL.md:41

    ssma-console instructs writing Oracle and SQL Server passwords in cleartext to ssma-variables.xml in the current folder (with Trust Server Certificate defaulting to true), displaying the resolved XML in chat for confirmation, and leaving the file on disk where it can be committed to git or logged.

    Fix: Redact secrets before displaying XML in chat, store ssma-variables.xml in a restricted temporary location deleted immediately after execution, and default trust-server-certificate to false.

  37. LowDouble-quoted echo of generated Markdown enables shell command substitution

    skills/vardoger-analyze/SKILL.md:111

    vardoger-analyze summarizes past Copilot CLI session transcripts into Markdown and instructs piping the result via echo "YOUR_PERSONALIZATION_HERE" | vardoger write. In Bash, backticks and $() inside double quotes execute as shell commands, so inline code formatting or prompt-injected text from session history can trigger command execution.

    Fix: Use a single-quoted heredoc (cat <<'EOF' | vardoger write --platform copilot --scope global) so the shell performs no command or variable expansion.

  38. LowPrivilege escalation to root via sudo npx inside untrusted workspace

    skills/sandbox-npm-install/scripts/install.sh:183

    When invoked with --playwright in a container where passwordless sudo is available, install.sh runs 'sudo npx playwright install-deps chromium' from inside $WORKSPACE_CLIENT after symlinking node_modules. npx resolves ./node_modules/.bin/playwright from the workspace, allowing a malicious dependency or postinstall script to execute arbitrary code as root.

    Fix: Run sudo commands from a root-owned directory outside the workspace using an explicit path to a trusted Playwright binary rather than resolving from workspace node_modules.

  39. LowUnpinned remote install scripts piped directly to shell

    skills/aspire/references/cli-reference.md:13

    The Aspire CLI setup instructions and deployment CI/devcontainer templates pipe remote scripts (https://aspire.dev/install.sh and install.ps1) directly into bash or PowerShell iex without version pinning or checksum verification, termasuk in CI workflows that hold cloud deployment credentials.

    Fix: Download the installer first, pin a specific release version, and verify its SHA-256 checksum before execution.

  40. LowPlaintext secret storage and over-broad permissions in Azure DevOps variable group example

    skills/azure-devops-cli/references/workflows-and-patterns.md:158

    Under Best Practices, the variable group automation snippet passes API_KEY=$API_KEY to 'az pipelines variable-group create --variables ... --authorize true'. This stores API_KEY as a non-secret plaintext variable visible to project readers and logs, exposes the value in process arguments, and grants every pipeline in the project access to the group.

    Fix: Create secret variables separately using 'az pipelines variable-group variable create --secret true' via environment variable input, and omit '--authorize true' so only specific pipelines are authorized.

  41. LowGenerated skill integration test disables all Copilot CLI safety prompts with --yolo

    skills/quality-playbook/references/review_protocols.md:564

    When generating RUN_INTEGRATION_TESTS.md for AI skills or LLM tools, quality-playbook includes a gh copilot command with the --yolo flag. Running Copilot CLI with --yolo disables all user confirmation gates for shell commands, file writes outside the workspace, and network access, allowing prompt injection inside a tested project to execute arbitrary commands.

    Fix: Remove --yolo from the template and scope tool permissions to the minimum required for the test directory.

  42. LowUnpinned third-party repository clone and install.sh execution

    skills/ponder/references/harness-notes.md:13

    The installation instructions clone an unpinned personal GitHub repository (Quinnod345/ponder) and immediately run install.sh, which links an executable into ~/.local/bin/ponder and installs skill symlinks across Claude Code and Codex. Without a pinned commit SHA or release verification, a compromise of that external repository compromises every agent harness on the machine; the same document also advises copying the codex binary to work around macOS Gatekeeper/amfid path checks.

    Fix: Pin the git clone to a reviewed commit SHA or signed tag, instruct users to inspect install.sh before running it, and remove advice to bypass macOS Gatekeeper checks.

  43. LowC# LSP configuration auto-downloads and runs unpinned prerelease packages at startup

    skills/lsp-setup/references/lsp-servers.md:65

    The C# language server snippet configures Copilot CLI to invoke dotnet dnx roslyn-language-server with --yes and --prerelease. Once saved to .github/lsp.json or user config, starting an LSP session on C# files can automatically download and execute unpinned prerelease packages from NuGet without confirmation.

    Fix: Pin roslyn-language-server to a specific vetted version instead of floating on --prerelease with --yes, and install the tool explicitly ahead of time.

  44. LowJava Streamable HTTP MCP server templates bind to all network interfaces without authentication

    skills/java-mcp-server-generator/references/streamable-http.md:41

    Unlike the Python MCP generator (which binds to 127.0.0.1 by default), the Java embedded Tomcat and Spring Boot templates listen on 0.0.0.0 with no authentication. In the Servlet example, DefaultServerTransportSecurityValidator checks Host: localhost:<port>, which stops browser DNS rebinding but still allows any attacker on the local network to call <machine-ip>:8080/mcp by sending a spoofed Host header and omitting Origin; the Spring Boot snippet omits securityValidator unless added manually from Gotchas.

    Fix: Bind embedded Tomcat and Spring Boot templates to 127.0.0.1 by default (for example via connector address or server.address=127.0.0.1) and include DefaultServerTransportSecurityValidator in the main Spring Boot example.

  45. LowMCP proxy setup runs unpinned npx -y mcp-remote and advises disabling browser shields

    skills/penpot-uiux-design/references/setup-troubleshooting.md:137

    The Claude Desktop configuration runs npx -y mcp-remote without a version pin (and with --allow-http) on every client launch, and the troubleshooting section recommends turning off Brave Shield on design.penpot.app to bypass local-network restrictions, weakening browser protections for that origin.

    Fix: Pin mcp-remote to an exact version in the MCP config and use narrower Private Network Access permissions instead of disabling browser shields globally for the site.

  46. LowBrowser Markdown snippet assigns unsanitized marked.parse output to innerHTML

    skills/markdown-to-html/references/marked.md:36

    Marked does not sanitize HTML in parsed Markdown, so assigning marked.parse(...) directly to element.innerHTML introduces cross-site scripting (XSS) whenever untrusted Markdown is rendered, and the example loads an unpinned script from cdn.jsdelivr.net without Subresource Integrity.

    Fix: Sanitize marked.parse output with DOMPurify before assigning to innerHTML, and pin the CDN script version with an integrity hash.

  47. LowCI workflow templates omit least-privilege permissions and use a mutable third-party action ref

    skills/repo-standardizer/templates/ci-rust.yml:13

    The Rust CI template references dtolnay/rust-toolchain@stable using a mutable branch ref instead of a pinned commit SHA, and all four CI workflow templates (Go, Node, Python, Rust) omit a top-level permissions: contents: read block, inheriting potentially broad GITHUB_TOKEN permissions.

    Fix: Pin third-party GitHub Actions to a full 40-character commit SHA and add permissions: { contents: read } to every CI workflow template.

  48. LowIncremental threat model creates persistent git worktree outside workspace using unvalidated commit SHA

    skills/threat-model-analyst/references/incremental-orchestrator.md:84

    Phase 0 reads baseline_sha from a prior report's 0-assessment.md file, runs git worktree add ../baseline-{baseline_sha_short} {baseline_sha} to create a checkout in the parent directory outside the project root, and forbids removing it in Rule I6. This writes persistent copies of repository code outside the workspace without asking the user, and passing an unvalidated baseline_sha string from markdown into shell commands risks git option or command injection when reviewing a repo that contains a crafted baseline report.

    Fix: Validate that baseline_sha and target_sha strictly match hexadecimal commit hashes ([0-9a-f]{7,40}), quote all shell arguments, and create temporary worktrees only inside a user-approved or gitignored directory with cleanup after analysis.

  49. LowStandalone verification harness uses a predictable shared /tmp directory

    skills/webmcpify/references/verify.md:128

    When verifying a project that does not already have Playwright installed, the webmcpify verification guide tells the agent to run 'mkdir -p /tmp/webmcpify-harness', install test packages there, and run './node_modules/.bin/playwright'. On a shared Linux or macOS machine, '/tmp' is writable by all local users and 'mkdir -p' succeeds even if another user already created that directory, allowing a local attacker to plant a malicious executable that runs under the victim's account.

    Fix: Create a private temporary directory with 'mktemp -d' instead of a fixed path under '/tmp'.

  50. LowCI workflow template clones and runs unpinned vcpkg bootstrap scripts

    skills/vcpkg/references/ci.md:110

    The multi-triplet GitHub Actions workflow fragment in the vcpkg skill clones 'https://github.com/microsoft/vcpkg' from its default branch without pinning a commit SHA or tag, then immediately executes 'bootstrap-vcpkg.sh' or 'bootstrap-vcpkg.bat'. If the upstream branch is compromised, CI pipelines using this template will execute untrusted code with access to workflow permissions and secrets.

    Fix: Check out a specific pinned commit SHA (matching the project's builtin-baseline) before running the bootstrap script.

  51. LowGeofeed HTML report loads third-party scripts from CDN without integrity hashes

    skills/geofeed-tuner/scripts/templates/index.html:10

    The generated geofeed HTML report loads Leaflet and h3-js JavaScript directly from unpkg.com without Subresource Integrity (integrity) attributes. If the CDN is compromised, malicious script could execute in the user's browser with access to all geofeed data rendered in the local report.

    Fix: Add sha384 Subresource Integrity hashes and crossorigin="anonymous" attributes to all external script and stylesheet tags, or bundle the libraries locally.

  52. LowWinMD cache generator falls back to wildcard NuGet package version

    skills/winmd-api-search/scripts/cache-generator/CacheGenerator.csproj:27

    When WinAppSdkVersion is not supplied, CacheGenerator.csproj restores Microsoft.WindowsAppSDK with Version="*", pulling whatever version is newest on NuGet at runtime. Although ExcludeAssets="all" prevents build targets or analyzers in that package from executing during compilation, unpinned wildcard resolution weakens supply-chain predictability.

    Fix: Pin a known-good fallback version of Microsoft.WindowsAppSDK instead of using Version="*".