outbound-destination-check — скилл для Claude Code
Скилл outbound-destination-check для Claude Code проверяет внешние сетевые соединения процессов и отличает штатный служебный трафик вендоров от активности вредоносного ПО.
Apache-2.0 ★ 71 обновлён 2026-11-15
npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a claude-codeЧто делает outbound-destination-check
Скилл определяет реальных адресатов исходящих сетевых запросов через прямой и обратный DNS-поиск с помощью системной команды getent. Проверка сопоставляет внешние IP-адреса со списками официальных хостов обновлений, зеркал пакетов, ACME-серверов Let's Encrypt и сервисов сбора телеметрии вроде Traefik или Grafana.
Процедура изучает флаги запуска процессов в ps или docker inspect, выявляя параметры отключения фоновых проверок в конфигурации сервиса. Анализ сопоставляет время сетевой активности с моментом запуска контейнера: периодические запросы через равные интервалы обычно вызваны штатными таймерами демона. Для работы скиллу требуются системные вызовы через host_exec.
Какие задачи решает
- определить владельца публичного IP-адреса перед изоляцией контейнера
- отличить регламентные опросы обновлений и телеметрии от связи с сервером злоумышленника
- проверить смещение времени сетевой активности относительно старта процесса
- найти параметры конфигурации для отключения нежелательных фоновых исходящих вызовов
Примеры запросов агенту
- Проверь адресата исходящего соединения для процесса с PID 1042
- Выясни, кому принадлежит публичный IP-адрес из алерта мониторинга сетевой активности
- Проанализируй сетевые вызовы веб-сервера и проверь настройки обновления
Как установить outbound-destination-check
1Через Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a claude-code2Вручную: положить файлы
~/.claude/skills/outbound-destination-check/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/outbound-destination-check/
1Через Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a codex2Вручную: положить файлы
.agents/skills/outbound-destination-check/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a cursor2Вручную: положить файлы
.cursor/skills/outbound-destination-check/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install Problem99-io/kodeblue blue/skills/outbound-destination-check2Вручную: положить файлы
.github/skills/outbound-destination-check/SKILL.md1Через Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a gemini-cli2Командой платформы
gemini skills install https://github.com/Problem99-io/kodeblue --path blue/skills/outbound-destination-checkВручную: положить файлы
.devin/skills/outbound-destination-check/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы об outbound-destination-check
Как установить outbound-destination-check в Claude Code?
Выполните команду npx skills add Problem99-io/kodeblue --skill outbound-destination-check -a claude-code в терминале.
Что делает outbound-destination-check?
Скилл выполняет прямой и обратный резолв хостов через getent, изучает аргументы запуска процесса и сопоставляет тайминги запросов с его стартом, чтобы исключить ложные срабатывания мониторинга.
Что нужно для работы outbound-destination-check?
Требуется поддержка вызовов host_exec, доступ к системным утилитам getent, ps, а также утилитам проверки пакетов или инспекции Docker.
Исходный текст
SKILL.md · Apache-2.0
--- name: outbound-destination-check description: Identify who an outbound connection really goes to before calling it hostile — reverse and forward name lookups, the software's own built-in phone-home (update checks, usage stats, ACME, package mirrors), and timing against the process start and earlier sightings. metadata: kodeblue-bundled: true --- # Who is this connection to? A service's own process connecting to a public address it was not seen using before is either an attacker's code calling home or the software calling its vendor. Many servers call home on their own: update checks, anonymous usage stats, plugin catalogs, certificate renewal. These calls are rare (once a day, or once some minutes after start), so a learning run often misses them and a `no_outbound` monitor fires on the first one. Find out which it is before you contain, because killing a service's main process takes the service down, and its supervisor starts it again. Do these checks in order. Each one costs one or two `host_exec` calls. ## 1. Name the destination - **Reverse name:** `host_exec ["getent","hosts","<ip>"]`. A vendor or CDN name (`lb-140-82-121-5-fra.github.com`) says who runs the address. A cloud name (`ec2-35-180-167-84.eu-west-3.compute.amazonaws.com`, `*.googleusercontent.com`, `*.cloudfront.net`) only names the cloud provider, not its customer. Thousands of vendors run there, so "AWS" or "a cloud IP" is neither hostile nor an answer: go on to the forward check. - **Forward match:** resolve the hostnames this software is known to contact (the table below, the software's own config, its documentation) with `host_exec ["getent","ahosts","<hostname>"]`. If the connection's exact IP is among the answers, you have named the service at that address. Resolve them now, not from memory: cloud addresses change. - No reverse name and no forward match leaves the destination unknown. That is not proof of an attack, but it is not an explanation either. ## 2. Look for the setting that makes the call Read how the process was started (`ps -o cmd -p <pid>`, `docker inspect <container>` for the arguments, its config file) and look for the switch that controls the call. A default that is on and not turned off is an explanation; name it, so a human can turn it off. | Software | Calls by default | When | Setting that controls it | |---|---|---|---| | Traefik | `update.traefik.io` (version check; it also reaches `api.github.com`), `collect.traefik.io` (usage stats, when enabled) | about 10 minutes after start, then every 24 hours | `--global.checknewversion=false`, `--global.sendanonymoususage=false` | | Traefik, Caddy, certbot, any ACME client | `acme-v02.api.letsencrypt.org`, `acme.zerossl.com` | at start and before certificates expire | the ACME / TLS settings | | Grafana | `grafana.com` (update check), `stats.grafana.org` (usage stats) | at start, then periodically | `check_for_updates`, `reporting_enabled` | | GitLab | `version.gitlab.com` | periodically | version check and Service Ping settings | | Coolify | `cdn.coollabs.io` (versions), its own instance (sentinel push) | periodically | auto-update settings | | Docker daemon | `registry-1.docker.io`, `auth.docker.io`, `production.cloudflare.docker.com` | on image pulls | registry mirrors | | Ubuntu | `motd.ubuntu.com`, `api.snapcraft.io`, `archive.ubuntu.com`, `security.ubuntu.com`, `esm.ubuntu.com`, `ntp.ubuntu.com` | timers (`systemctl list-timers`) | `motd-news`, snapd, apt and timesyncd config | The table holds examples, not an allow list. A match explains the destination only when the process making the call is the one in the table, its binary is the packaged or image one (`sha256sum`, `dpkg -S`, the image's own file times), and the call happened when the software makes it. ## 3. Compare the timing with the process start - Get the start time: `host_exec ["ps","-o","pid,lstart,etime,cmd","-p","<pid>"]`. - Work out the offset of each sighting from that start. Built-in timers show up as a fixed offset after start (10 minutes), a fixed period (12 or 24 hours, so a sighting at start + N days to the minute), or both. - Earlier sightings: `findings_list` for the same monitor and earlier incidents with the same title. The same destinations at the same offset after several starts are a timer in the software. An implant does not reset its schedule to the process start. ## 4. Your own kill can cause the repeat When an earlier session killed the process, the supervisor started a new one, and the timer runs again from that start. A sighting the same few minutes after every kill is the new process doing its normal start-up call, not the attacker coming back. Check the offset from the latest start before you call a repeat "re-compromise", and never kill the same service again for the same recurring finding without new evidence (a child process, a dropped or changed file, a destination that does not resolve to the vendor). ## Verdict - **Explained:** the destination resolves to the software's own endpoint, the setting is on, the binary is the image's or package's own, and the timing fits. The verdict is `false_positive` (the monitor did not know about the call). Do not contain. In the report, name the endpoint, the setting a human can turn off, and the timing you saw. Record the call in the `Server role and services` definition with `patch_definition`. Do not add the IP to the monitor's `allow_remotes`: vendor addresses on shared clouds change and are shared with other customers; turning the call off is the fix. - **Not explained:** no name, a name unrelated to the software, a destination that matches none of its endpoints, timing unrelated to the start, or other signs (a child process, a dropped file, a changed binary). Treat it as the attack explanation and contain as the responder procedure says. - **Partly explained** (the name fits but something else does not): say what fits and what does not, and resolve `inconclusive` rather than killing a service on a guess.
© Problem99-io · Apache-2.0 · Problem99-io/kodeblue