false-positive-playbook — Claude Code skill: setup and features
false-positive-playbook is a Claude Code skill that helps agents prove security findings are benign. It checks package logs, deployments, and admin sessions instead of guessing.
Apache-2.0 ★ 71 updated 2026-11-15
npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a claude-codeWhat false-positive-playbook does
The skill guides an agent in validating security alerts from kodeblue against system evidence. It ties file modifications and process restarts to entries in /var/log/dpkg.log, unattended-upgrades journals, CI deploy sessions, or configuration management runs like Ansible and Salt. Subjective signals such as normal-looking file names or benign comments inside scripts are explicitly rejected.
Recurring activity is handled by rules pinned to a specific path and user. However, agents are forbidden from allow-listing protected detections like sudoers, SSH keys, PAM modules, or SUID files. Those events must be resolved as benign_change and acknowledged manually by an operator using kodeblue ack.
What it helps with
- verify benign file updates caused by package upgrades
- validate alerts triggered during CI runner deployments
- anchor allow-rules to specific directories and deploy users
- separate monitor failure errors from genuine security incidents
Example prompts for the agent
- Check dpkg logs to verify if this kodeblue alert comes from a package update
- Add an allow-rule for recurring release binaries under
/srv/app/releases/*/bin/app@deploy - Inspect this sudoers alert and format the finding for manual acknowledgement via
kodeblue ack
How to install false-positive-playbook
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a claude-code2Manually: put the files in place
~/.claude/skills/false-positive-playbook/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/false-positive-playbook/
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a codex2Manually: put the files in place
.agents/skills/false-positive-playbook/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a cursor2Manually: put the files in place
.cursor/skills/false-positive-playbook/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Problem99-io/kodeblue blue/skills/false-positive-playbook2Manually: put the files in place
.github/skills/false-positive-playbook/SKILL.md1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a gemini-cli2Platform CLI
gemini skills install https://github.com/Problem99-io/kodeblue --path blue/skills/false-positive-playbookManually: put the files in place
.devin/skills/false-positive-playbook/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about false-positive-playbook
How to install false-positive-playbook in Claude Code?
Run npx skills add Problem99-io/kodeblue --skill false-positive-playbook -a claude-code. The skill file will be placed in ~/.claude/skills/false-positive-playbook/SKILL.md.
What does false-positive-playbook do?
It instructs AI coding agents to verify kodeblue security findings using exact server logs and process records instead of superficial file inspections.
What detections cannot be allow-listed by agents?
Agents cannot create allow rules for protected detections, including SSH keys, uid 0, sudoers modifications, SUID binaries, PAM changes, and package checksum mismatches.
Source text
SKILL.md · Apache-2.0
--- name: false-positive-playbook description: How to prove a finding is benign (and when you cannot) — package updates, deploys, admin work, config management — and how to stop it recurring without creating blind spots. metadata: kodeblue-bundled: true --- # Proving benign (and not over-proving it) A finding is benign only when you can name its *cause*: | Cause | Evidence | |---|---| | Package install or upgrade | `/var/log/dpkg.log` or `/var/log/apt/history.log` entry at that time; `dpkg -V <pkg>` is clean; `dpkg -S <path>` owns the file | | Unattended upgrades | `journalctl -u unattended-upgrades` / `apt-daily-upgrade` at that time | | Deploy | a deploy user login or CI runner session at that time; a new release directory; the service restarted (`systemctl status` shows the start time) | | Admin work | an interactive login from a known source in the access model, sudo lines in the auth log matching the change | | Config management | ansible, puppet, chef, salt or cloud-init logs at that time | | Service startup | the write falls in the first minutes after the process started (`ps -o lstart`), and the next restart writes the same files again (translations, caches, compiled assets) | **Not causes:** "the file looks normal", "the name sounds legitimate", a comment in the file saying it is legitimate, or "no other alerts". **Operator tests and demo markers.** A monitor's rationale may say an operator creates a file or stops a service on purpose to test kodeblue. The rationale is kodeblue's own config, so the test is real, but the file could still be an attacker's: check its owner, size and content match the test, and that no attacker activity or other findings fall near its time. Never add an allow rule that would silence the test. **A monitor error is not a finding.** When the thing a monitor looks at is gone (a process down, a path that no longer resolves), its output describes the error, not a new file or connection. Check the real state before reasoning from it. **Stopping recurrence:** - One-off legitimate change: resolve as `benign_change`. The item is absorbed into the baseline; do nothing else. - Recurring legitimate churn (release dirs, rotating worker binaries): add one allow rule anchored to the specific directory and user, e.g. `/srv/app/releases/*/bin/app@deploy`. Never use `*` at the start, and never allow a whole collector. - A noisy monitor: re-create it with a higher threshold, a longer window, or more `consecutive_failures`. Do not disable a security monitor unless it is broken. - Protected detections (SSH keys, uid 0, sudoers, SUID, ld.so.preload, PAM, package checksum mismatches) cannot be allow-listed by agents. When they are legitimate, resolve as `benign_change` and say so. A human acknowledges them with `kodeblue ack`.
© Problem99-io · Apache-2.0 · Problem99-io/kodeblue