← MCP server and skill grades

k-skill

NomaDamas/k-skill

D
Type
Plugin
Scanned
Nov 13, 2026
Version
9766d20fb1fe
Files read
2169

This plugin gives AI coding agents over 130 skills to access Korean public data, shopping, transit bookings, financial and legal services, and HWP document tools. The biggest risk is that skills automatically download and run unpinned online packages at runtime, one skill lets a remote server override local rules to run commands, and other flaws disable HTTPS security checks on bus-booking pages or let scripts run from outside the skill folder.

What it can reach

Files
Reads and writes local documents (such as HWP, PDF, CSV, JSON, and HTML files), cache and settings folders (~/.cache/k-skill and ~/.config/k-skill), global AI skill folders (~/.claude/skills and ~/.agents/skills), coding-agent session logs, and macOS KakaoTalk chat archives.
Network
Connects to the project's proxy server (k-skill-proxy.nomadamas.org), dozens of Korean government and commercial websites and APIs, third-party remote MCP and HeyYumi servers, code registries (npm, PyPI, GitHub), and PostHog for default-on usage analytics.
Shell commands
Runs local terminal commands, bundled Python and Node scripts, git and curl requests, and npx or pip commands that download and run software packages on the fly.
Secrets
Reads API keys, account logins, OAuth tokens, and session cookies from environment variables, credential vaults, ~/.config/k-skill/secrets.env, and ~/.config/heyyumi/credentials.json for Korean government, financial, and commercial services.
Browser
Connects to open local browsers (such as Chrome, BrowserOS, Aside Browser, and CloakBrowser) to read and interact with logged-in pages, and launches automated Chromium browsers for logins, searches, and bookings.
System
Can set up a daily background task (cron job or Windows Scheduled Task) to check for skill updates after user approval.

Findings

4 high15 medium14 low1 info
  1. HighBus booking helpers disable HTTPS certificate verification

    express-bus-booking/scripts/kobus_express_booking.py:71

    Both express-bus-booking/scripts/kobus_express_booking.py and intercity-bus-booking/scripts/intercity_bus_search.py unconditionally disable TLS certificate verification using ssl._create_unverified_context(), and the KOBUS helper also lowers OpenSSL cipher security to DEFAULT@SECLEVEL=1. Anyone on the same Wi-Fi or network path can intercept the booking traffic without a valid certificate, alter schedules or seat holds, and inject malicious phishing forms or scripts into the checkout and card-entry HTML files saved to disk.

    Fix: Use ssl.create_default_context() so server certificates are verified by default, and do not silently disable TLS verification on booking or checkout flows.

  2. HighPath traversal in k-skill CLI skill name escapes bundled skills directory

    packages/k-skill-cli/src/assemble.js:42

    While resolveBundledAsset() blocks '../' in the second argument (the asset path), loadSkill() joins the first argument (skillName) directly into skillsDir without checking for path separators or verifying that the resolved directory stays inside skillsDir. If prompt injection tricks the agent into passing a relative path such as '../../../../path/to/untrusted-dir' as the skill name, 'k-skill exec' will execute arbitrary scripts from any local directory containing a minimal skill.json and scripts/ subfolder, bypassing the CLI confinement boundary.

    Fix: Validate that skillName matches /^[a-z0-9-]+$/ and verify with fs.realpathSync that the resolved skill directory is strictly inside skillsDir before loading skill.json or assets.

  3. HighHeyYumi skill lets remote server responses override local rules and specify CLI commands

    heyyumi/instruction.md:38

    The heyyumi skill instructs the coding agent on every task to fetch live rules from the third-party HeyYumi server (npx -y heyyumi guide), explicitly stating that if the server output differs from the local instruction file, the agent must follow the remote guide instead, and must execute CLI commands returned in the server output's commands and cli.next fields. If the remote HeyYumi server or its response is compromised, it can override local safety boundaries (such as user approval gates) and direct the coding agent to run arbitrary commands on the user's machine.

    Fix: Treat all remote guide and API responses strictly as untrusted data that cannot override local skill rules, and restrict execution to a fixed allowlist of known heyyumi subcommands and flags rather than running server-supplied commands or cli.next strings.

  4. HighHeyYumi skill lets remote server instructions override local rules and executes server-returned commands

    packages/k-skill-cli/skills/heyyumi/instruction.md:38

    The heyyumi skill instructs the agent to fetch live operating rules from the external HeyYumi server on every task (npx -y heyyumi guide) and explicitly states that if those remote rules differ from the local skill document, the agent must follow the remote server's rules instead. It also directs the agent to execute CLI commands mapped in the server's commands output and to run any command returned in cli.next. If the third-party HeyYumi server or its response data is compromised, it can bypass local approval guardrails (such as clarify) and steer the agent into running arbitrary commands on the user's machine.

    Fix: Treat all output from heyyumi and its remote server strictly as untrusted data, never allow remote guide text to override local safety rules, and validate any cli.next follow-up against a strict allowlist of heyyumi subcommands and flags.

  5. MediumEvery skill stub downloads and runs unpinned npm code and instructions at runtime

    k-skill-setup/SKILL.md:20

    All generated SKILL.md stubs tell the coding agent to run 'npx -y @nomadamas/k-skill@0 instruct <skill>' and treat its output as primary instructions, while packages/k-skill-cli/src/update.js runs 'npx --yes skills add NomaDamas/k-skill --all -g'. Because '@0' is a floating major-version range rather than a pinned version or hash, normal skill use pulls and executes the latest 0.x release from npm at runtime; a compromise of the npm package or repository would automatically deliver arbitrary code and agent instructions to existing installations.

    Fix: Bundle the static instruction.md and helper scripts directly in the installed plugin or pin exact package versions and integrity hashes instead of using a floating '@0' range with npx -y.

  6. MediumPopbill CLI generic call subcommand bypasses mutation and approval guards

    popbill/scripts/popbill_cli.py:179

    The 'call' subcommand checks guard_dangerous() using a prefix denylist (popbill_safety.py) and then invokes getattr(service, args.method) without blocking underscore-prefixed SDK methods. If an agent is tricked via prompt injection when Popbill credentials are configured, it can call internal SDK methods like '_getToken' (printing live bearer session tokens into chat) or '_httppost', as well as public mutating methods not in the prefix list such as 'reverseIssue' or 'tempSave', without passing --yes-i-understand or --allow-production.

    Fix: Reject any method name starting with '_' and replace the prefix denylist in popbill_safety.py with an explicit allowlist of known read-only SDK methods, requiring explicit approval flags for everything else.

  7. MediumUnvalidated URLs in lookup helpers allow local network SSRF and file:// reads

    daangn-realty-search/scripts/daangn_realty.py:45

    Unlike naver-blog-research and company-analysis, which restrict URLs to expected HTTPS hosts, several read-only lookup scripts—including daangn_realty.py, daangn_jobs.py, daangn_cars.py, daangn_used_goods.py, ohou_today_deal.py, and geeknews_search.py—pass user- or agent-supplied URLs straight to Python's urllib.request.urlopen(). If prompt injection tricks the agent into supplying a localhost http:// URL (such as local browser CDP ports) or a file:// URI, these domain-specific lookup helpers can read internal endpoints or local files.

    Fix: Validate in each helper that parsed URLs use the https:// scheme and match the expected service hostname (such as realty.daangn.com, www.daangn.com, or ohou.se) before calling urlopen().

  8. MediumTLS certificate verification is unconditionally disabled in KOBUS booking helper

    packages/k-skill-cli/skills/express-bus-booking/scripts/kobus_express_booking.py:71

    The KOBUS express-bus helper creates an unverified SSL context (ssl._create_unverified_context()) and lowers OpenSSL cipher security to SECLEVEL=1 for all requests to www.kobus.co.kr. An attacker on the same Wi-Fi or network path can intercept timetable and seat-hold requests without any TLS error, steal temporary seat-hold tokens, and inject malicious HTML or scripts into the saved kobus-checkout-response.html file.

    Fix: Use ssl.create_default_context() so server certificates are verified by default, and only allow reduced cipher security if needed while keeping certificate and hostname verification enabled.

  9. MediumTLS certificate verification is unconditionally disabled in Tmoney intercity bus helper

    packages/k-skill-cli/skills/intercity-bus-booking/scripts/intercity_bus_search.py:72

    The Tmoney intercity-bus helper unconditionally disables HTTPS certificate verification via ssl._create_unverified_context(). During seat holding, it fetches the card-information entry page over this unverified connection and saves the raw HTML to tmoney-intercity-card-info.html, allowing a network man-in-the-middle attacker to hijack held seats or serve a spoofed card-input page.

    Fix: Verify server certificates using ssl.create_default_context() by default instead of disabling certificate verification.

  10. MediumPath traversal in LCK analytics report writers allows overwriting arbitrary JSON files

    packages/k-skill-cli/skills/lck-analytics/scripts/analyze-live-game.js:37

    analyze-live-game.js builds the output path directly from --game without checking for path separators or parent-directory segments (and build-match-report.js does the same with --team). If prompt injection tricks the agent into passing a value such as x/../../../../.mcp, path.join resolves outside the cache directory and overwrites arbitrary .json files on the user's machine.

    Fix: Validate that gameId and team arguments match a strict safe pattern (such as /^[A-Za-z0-9_-]+$/) and verify that path.resolve(reportFile) stays inside paths.reports before writing.

  11. MediumUnvalidated detail URL in Daangn helpers allows SSRF and local JSON file disclosure

    packages/k-skill-cli/skills/daangn-cars-search/scripts/daangn_cars.py:65

    The detail subcommand in daangn_cars.py (as well as daangn_jobs.py, daangn_used_goods.py, and daangn_realty.py) passes args.url to urllib.request.urlopen without checking the URL scheme or hostname. If an agent is tricked into passing a URL ending with # (which strips the appended query suffix in urllib.request), such as file:///path/to/config.json# or http://127.0.0.1:9222/json/version#, the script reads and prints the local JSON file or internal HTTP response.

    Fix: Parse args.url with urllib.parse.urlparse, require the scheme to be https and the hostname to be www.daangn.com or realty.daangn.com, and reject URLs containing fragments or userinfo.

  12. MediumFlight search helper silently installs Python packages from PyPI at runtime without hashes

    packages/k-skill-cli/skills/flight-ticket-search/scripts/flight_ticket_search.py:24

    When invoked, flight_ticket_search.py automatically creates a virtual environment under ~/.cache/k-skill/flight-ticket-search/venv and runs pip install -q fast-flights==2.2 before re-executing itself. Although fast-flights is version-pinned, pip installs all of its transitive dependencies from PyPI at runtime without hash verification or user confirmation.

    Fix: Pin all direct and transitive dependencies with --require-hashes in a requirements lockfile, or document and require explicit setup before installing third-party packages.

  13. MediumMultiple skills download and execute unpinned npm packages at runtime without confirmation

    heyyumi/instruction.md:26

    Several default-enabled skills instruct the agent to run third-party npm packages with npx -y / npx --yes or npm install -g without pinning an exact version—and heyyumi explicitly forbids version ranges so npx always pulls new releases. Similar unpinned execution appears in bunjang-search (npx --yes bunjang-cli), olive-young-search and korean-cinema-search (npx --yes daiso and unpinned git clone), hwp and gov-overseas-trip-report (npx --yes --package kordoc --package pdfjs-dist), kbo-results (npm install -g kbo-game), k-skill-setup (npx --yes skills), and every SKILL.md stub (floating @nomadamas/k-skill@0). If any of those upstream npm accounts is compromised, invoking the skill immediately runs attacker code on the user's computer.

    Fix: Pin exact package versions (and commit SHAs for git clones, as iros-registry-automation already does) or bundle vetted dependencies inside workspace packages.

  14. MediumPredictable shared /tmp directories used for npm execution and cached data

    gov-overseas-trip-report/instruction.md:171

    The fallback workflow in gov-overseas-trip-report creates a fixed /tmp/kordoc-run directory, changes into it, installs npm packages, and runs npx kordoc, while k-dart caches and extracts /tmp/dart_corp.zip into /tmp/dart_corp. On a shared multi-user machine, another local user can pre-create /tmp/kordoc-run with a malicious node_modules/.bin/kordoc or package.json (or plant symlinks/poisoned XML in /tmp/dart_corp) to execute arbitrary code or tamper with data when the skill runs.

    Fix: Create private per-run directories with mktemp -d and 0700 permissions instead of hardcoded paths under /tmp.

  15. MediumSkills automatically run unpinned npm self-updates and overwrite local skill installs

    packages/k-skill-cli/templates/core.md:6

    Every SKILL.md stub delegates its primary instructions at runtime to npx -y @nomadamas/k-skill@0 instruct <skill>, and the shared core.md template (active in all modes) instructs the agent to run npx -y @nomadamas/k-skill@0 update before using CLI tools. Because @0 floats across all 0.x.y npm releases and -y skips confirmation, invoking a simple read-only lookup skill causes the agent to download the latest npm release and overwrite coding-agent skill directories (such as ~/.agents/skills) without asking the user, bypassing any git commit or version review performed at install time.

    Fix: Pin @nomadamas/k-skill to an exact audited version in generated SKILL.md stubs, and remove the automatic pre-execution update rule from core.md so updates only run when explicitly requested by the user.

  16. MediumUnpinned third-party bunjang-cli package downloaded and executed at runtime

    packages/k-skill-cli/skills/bunjang-search/instruction.md:16

    The bunjang-search skill instructs the agent to run npx --yes bunjang-cli with no version pin and --yes auto-confirmation for searching, viewing items, and logging into Bunjang accounts. If the third-party bunjang-cli npm package or its publisher account is compromised, invoking this skill will automatically download and execute arbitrary code on the user's machine with access to local files and Bunjang login sessions.

    Fix: Pin bunjang-cli to a specific audited version (for example npx --yes bunjang-cli@<exact-version>) or bundle a lockfile-pinned dependency.

  17. MediumExternal merchant names interpolated into single-quoted shell command template

    packages/k-skill-cli/skills/bccard-eatpl-search/instruction.md:58

    Step 5 of bccard-eatpl-search tells the agent to take merNm (merchant name) and addr (address) from the external eat.pl search response as-is and pass them into a single-quoted bash command. Merchant names frequently contain single quotes, and if an upstream listing or manipulated response contains a single quote followed by shell metacharacters or prompt-injection text that the agent substitutes verbatim, it can break out of the single quotes and execute arbitrary shell commands.

    Fix: Pass merchant records via stdin or a JSON file to kakao-place-rating.mjs instead of interpolating external API fields into shell command templates, and instruct the agent to escape shell arguments safely.

  18. MediumMultiple skills download and execute unpinned third-party packages and repositories at runtime

    packages/k-skill-cli/skills/heyyumi/instruction.md:26

    Several skills instruct the agent to download and run external third-party code without version or commit pins, using -y/--yes to skip confirmation. In heyyumi, the instructions explicitly forbid adding a version range to npx -y heyyumi so that new upstream releases run immediately; similarly, olive-young-search and korean-cinema-search run unpinned npx --yes daiso or clone https://github.com/hmmhmmhm/daiso-mcp.git at HEAD and run npm install, and hwp and gov-overseas-trip-report run unpinned npx --yes --package kordoc --package pdfjs-dist (while every SKILL.md stub also pulls instructions from floating @nomadamas/k-skill@0). If any of those upstream npm accounts or repositories is compromised, a normal lookup will automatically execute malicious code on the user's computer.

    Fix: Pin third-party npm packages to exact reviewed versions and pin fallback Git clones to a reviewed 40-character commit SHA, matching the scripts/upstream.pin pattern already used in iros-registry-automation, korean-jangbu-for, and korean-privacy-terms.

  19. MediumFixed /tmp directories used for npm install and archive extraction allow local hijacking

    packages/k-skill-cli/skills/gov-overseas-trip-report/instruction.md:171

    In gov-overseas-trip-report, the fallback workflow creates a predictable /tmp/kordoc-run directory, runs npm install kordoc pdfjs-dist inside it, and executes npx kordoc /tmp/report.pdf. On a shared multi-user machine, another local user can pre-create /tmp/kordoc-run with a malicious package.json or node_modules/.bin/kordoc (or plant symlinks at /tmp/report.pdf or /tmp/dart_corp.zip and /tmp/dart_corp/CORPCODE.xml in k-dart) to hijack execution or poison data when the victim's agent runs the skill.

    Fix: Use mktemp -d with mode 0700 for temporary working directories and downloaded files (matching the safe pattern already used in iros-registry-automation/instruction.md) instead of hardcoded paths under /tmp.

  20. LowAPI keys transmitted over unencrypted HTTP to Korean Open Data endpoints

    fine-dust-location/scripts/fine_dust.py:14

    When direct API mode is used (KSKILL_PROXY_BASE_URL disabled), fine_dust.py sends the user's AIR_KOREA_OPEN_API_KEY in the query string over plaintext http://apis.data.go.kr URLs instead of https://. Several server-side adapters in packages/k-skill-proxy (airkorea.js, lh-notice.js, and Seoul Open Data routes in server.js) similarly transmit API keys over plaintext http://, exposing those keys to passive network eavesdroppers.

    Fix: Switch http://apis.data.go.kr URLs in fine_dust.py, airkorea.js, lh-notice.js, and server.js to https://apis.data.go.kr, and use HTTPS endpoints wherever supported.

  21. LowCLI sends default-on usage telemetry with a persistent user ID

    packages/k-skill-cli/src/analytics.js:87

    Every k-skill CLI invocation generates and stores a persistent UUID in ~/.config/k-skill/analytics-id and posts the invoked command and skill name to PostHog (https://us.i.posthog.com) unless KSKILL_ANALYTICS_DISABLED=1 is set. Although arguments and file contents are not sent and the README mentions this opt-out, recording which specific skills a user runs (including medical, legal, tax, and financial skills) under a persistent identifier leaks usage patterns without asking for consent first.

    Fix: Make analytics opt-in, or avoid persisting a cross-session identifier when reporting aggregate command counts.

  22. LowAir Korea direct API fallback sends API key and location over unencrypted HTTP

    packages/k-skill-cli/skills/fine-dust-location/scripts/fine_dust.py:14

    Both Air Korea API base URLs are hardcoded with http:// instead of https://. When proxy mode is disabled (KSKILL_PROXY_BASE_URL=off) or bypassed with --station-file, the script sends the user's AIR_KOREA_OPEN_API_KEY and converted location coordinates in the URL query string in cleartext over the network.

    Fix: Change STATION_SERVICE_URL and MEASUREMENT_SERVICE_URL from http://apis.data.go.kr/... to https://apis.data.go.kr/...

  23. LowSubstring host check and redirects allow SSRF bypass in overseas-trip surface discovery

    packages/k-skill-cli/skills/gov-overseas-trip-report/scripts/gov_overseas_trip_report.py:977

    In discover --urls, the allowlist checks urllib.parse.urlparse(url).netloc with 'open.go.kr' in host and does not restrict the URL scheme or HTTP redirects. An agent tricked into passing a URL such as http://open.go.kr.127.0.0.1.nip.io:8080/ bypasses the check, allowing requests to internal HTTP services and returning the page title and final URL.

    Fix: Require parsed.scheme == 'https', inspect parsed.hostname (matching exact open.go.kr or .go.kr / .or.kr suffixes), block private/loopback resolved IPs, and disable or re-validate redirects.

  24. LowKAMIS dry-run mode fails to redact URL-encoded API keys

    packages/k-skill-cli/skills/kamis-food-price/scripts/run_kamis.py:162

    When --direct and --dry-run are used together, run_kamis.py builds the URL with urllib.parse.urlencode(params) and then attempts to scrub the secret using url.replace(key, '<redacted>'). If KSKILL_KAMIS_API_KEY contains any character that urlencode percent-escapes (such as =, +, or /), the replacement misses and prints the URL-encoded secret to stdout.

    Fix: Build the dry-run URL directly from display_params (with p_cert_key already set to '<redacted>') before calling urllib.parse.urlencode, without resolving or interpolating the real secret.

  25. LowPopbill CLI mutation and production safety gate can be bypassed via private or unprefixed methods

    packages/k-skill-cli/skills/popbill/scripts/popbill_cli.py:179

    The popbill skill documents that mutating or billable calls require --yes-i-understand and production mutations additionally require --no-test --allow-production. However, command_call dispatches arbitrary attributes on the SDK service instance via getattr(service, args.method) without blocking underscore-prefixed helpers (such as _httppost or _httppost_files) and relies on a prefix denylist in popbill_safety.py that misses mutating SDK methods like reserveCancel, reserveCancelRN, and saveMemo, as well as token/SSO URL methods like getToken and getAccessURL. A prompt-injected agent can invoke these methods in production without passing the required approval flags.

    Fix: Reject method names starting with an underscore in command_call and replace the prefix denylist in popbill_safety.py with an explicit allowlist of verified read-only SDK methods.

  26. LowPlaintext API keys passed in curl command-line arguments and secrets.env sourced as shell code

    korean-transit-route/instruction.md:48

    Although every SKILL.md stub states never to place plaintext credentials in shell arguments, korean-transit-route and k-dart instruct the agent to pass ODSAY_API_KEY and API_K_DART directly in curl URLs and --data-urlencode arguments, exposing the expanded keys in the OS process table (ps) to other local users and potentially in agent command logs. In addition, korean-transit-route sources ~/.config/k-skill/secrets.env with . in bash and set -a, executing any shell syntax in the file and exporting every stored secret to child processes.

    Fix: Perform authenticated HTTP requests inside a Python or Node helper that parses key-value pairs from secrets.env without shell evaluation and reads API keys from memory rather than command-line arguments.

  27. LowSkills documented as strictly read-only attach mutating action profiles in skill.json

    toss-investment/skill.json:6

    Several skills whose README entries and instruction files explicitly promise read-only behavior—including toss-investment (lookup-only brokerage account/holdings API), naver-ad-performance (read-only ad stats), ticket-availability (no booking/payment due to anti-macro law), flight-ticket-search, and ohou-today-deal—include action:account, action:booking, or action:commerce in skill.json. In runtimes where action templates are enabled, the CLI injects instructions encouraging the agent to proceed past lookup into account, booking, or commerce actions despite the documented read-only guarantee.

    Fix: Replace action:* profiles with lookup in skill.json for all skills that are documented and intended to be strictly read-only.

  28. LowGeneric credential template tells agent to request missing secrets through host input

    packages/k-skill-cli/templates/vault.md:4

    Although each SKILL.md states never to ask for or store plaintext credentials in chat or files, the generic-mode vault.md template instructs the agent to request missing credentials through the safest input surface the host provides and save them to ~/.config/k-skill/secrets.env. In standard coding agents such as Claude Code, Cursor, and Codex, there is no isolated secret prompt, so asking the user for a credential sends the plaintext secret into the chat transcript and to the remote LLM provider.

    Fix: In generic mode, instruct the agent never to collect secrets via chat or prompt tools, and instead guide the user to set the variable manually in their own terminal or ~/.config/k-skill/secrets.env.

  29. LowSkill instructions forbid disclosing proxy and server details in user responses

    packages/k-skill-cli/skills/bccard-eatpl-search/instruction.md:11

    The bccard-eatpl-search skill sends location and restaurant queries to k-skill-proxy.nomadamas.org (and potentially an operator relay), which is not mentioned in its SKILL.md or README table entry, while its instructions strictly forbid ever mentioning endpoints, k-skill-proxy, relays, or servers in user-visible answers. While intended to keep restaurant tables free of developer jargon, an unconditional prohibition prevents the agent from answering transparently if the user asks where their data was sent.

    Fix: Limit the jargon restriction to default recommendation output and explicitly permit the agent to disclose endpoints, proxy servers, and error codes when the user asks about network behavior or troubleshooting.

  30. LowSetup workflow instructs the agent to run an automatic global --all update before using the CLI

    packages/k-skill-cli/skills/k-skill-setup/instruction.md:22

    Although k-skill-setup states at the top that external state changes require prior user consent, Section 2 instructs the agent to run npx -y @nomadamas/k-skill@0 update before using the CLI. That command runs npx --yes skills add NomaDamas/k-skill --all -g, which silently installs or overwrites all 134 skills globally across every coding agent on the machine—even if the user intentionally installed only a single skill (--skill <name>) or removed unwanted skills with k-skill-cleaner.

    Fix: Default pre-flight checks to npx -y @nomadamas/k-skill@0 update --check and require explicit user confirmation before running update or installing --all -g.

  31. LowAPI keys are passed directly in curl command-line arguments

    packages/k-skill-cli/skills/korean-transit-route/instruction.md:48

    Despite the hard rule in every SKILL.md never to pass plaintext credentials in shell arguments, korean-transit-route passes ODSAY_API_KEY (${KEY}) directly in the curl URL argument, and k-dart similarly passes $API_K_DART in curl URL and --data-urlencode arguments. Shell variable expansion exposes these secrets in the operating system process list (ps) to other local processes and can record them in agent command logs.

    Fix: Read API keys inside a Python or Node helper from os.environ or pass query/header secrets to curl via stdin (curl --config -) so plaintext keys never appear in command-line arguments.

  32. LowFeature guides instruct running unpinned third-party CLIs and live external server rules

    docs/features/heyyumi.md:58

    Several feature guides (including heyyumi.md, bunjang-search.md, olive-young-search.md, and korean-cinema-search.md) instruct running unpinned third-party npm packages with npx -y (such as heyyumi, bunjang-cli, and daiso) or cloning unpinned git HEADs, and heyyumi.md explicitly discourages version pinning while fetching live agent rules from an external server via heyyumi guide. If a user or agent follows these guides, a compromised third-party package or server update can run unreviewed code or alter agent behavior.

    Fix: Pin third-party npm packages and git repositories to reviewed exact versions or commit SHAs, matching the upstream.pin pattern already used in iros-registry-automation and korean-jangbu-for.

  33. LowDirect fallback example sends API key over unencrypted HTTP and in CLI arguments

    docs/features/fine-dust-location.md:85

    The direct AirKorea fallback example calls http://apis.data.go.kr over cleartext HTTP while passing AIR_KOREA_OPEN_API_KEY in curl command-line arguments. Anyone following this fallback exposes their API key to network eavesdroppers in transit and to other local users via process listings (similar CLI-argument key passing appears in k-dart.md and korean-transit-route.md).

    Fix: Use https://apis.data.go.kr instead of http:// and pass secrets via curl config stdin (--config -) rather than command-line arguments.

  34. InfoOpt-out CLI usage analytics enabled by default

    packages/k-skill-cli/src/analytics-policy.json:17

    The CLI sends anonymous invocation telemetry (command, skill name, CLI version, runtime mode, and success status) to PostHog by default unless KSKILL_ANALYTICS_DISABLED=1 is set in the environment. Command arguments, raw inputs, and file contents are excluded by policy, and this behavior is documented in README.md.

    Fix: Set KSKILL_ANALYTICS_DISABLED=1 in your environment if you do not want CLI usage statistics sent externally.