Aikido triage

Aikido Findings Triage Workflow

You are triaging an Aikido SAST/SCA/secret-scanning CSV export against a local codebase. Your job is to read every finding, investigate the flagged code, and verdict each one with evidence. At the end you produce a reviewed CSV and a self-contained HTML report.

Do not guess. Read the actual files before rendering a verdict.

---

Arguments

Parse from the user's invocation:

• CSV_PATH — path to the Aikido CSV export
• CODEBASE_PATH — absolute path to the local codebase to investigate

If either argument is missing, ask the user before proceeding.

---

Phase 1 — Parse the CSV

1. Read CSV_PATH with the Read tool.

2. Parse every row. Key columns to extract per finding:

   • aikido_issue_id — unique ID
   • type — sast | open_source | leaked_secret | eol
   • severity — critical / high / medium / low
   • affected_file — relative path to the flagged file
   • related_cve — CVE or advisory ID (may be empty)
   • rule — rule name that fired
   • start_line / end_line — flagged line range
   • installed_version / patched_version — for SCA findings
   • affected_package — for SCA/EOL findings

3. Group findings by type so you can batch-investigate efficiently.

---

Phase 2 — Investigate Each Finding

Work through findings type by type. For each one, read the flagged file and apply the investigation playbook for that type.

2a — leaked_secret findings

Step 1 — Does the file exist?

Use the Read tool on CODEBASE_PATH/affected_file. If the file does not exist: → Verdict: CLOSE — File Removed → Note: "File does not exist at HEAD. Finding references removed code in git history." → Move on. Do not investigate further.

Step 2 — Read the flagged lines.

Read start_line to end_line (expand by ±3 lines for context).

Apply these patterns:

What you see	Verdict
ENV.fetch(...), ENV[...], ENV.fetch(..., nil)	CLOSE — False Positive (env var read)
${{ secrets.* }} (GitHub Actions)	CLOSE — False Positive (GH Actions secret)
${VARIABLE_NAME} in .npmrc/.env.example	CLOSE — False Positive (env var placeholder)
--mount=type=secret in Dockerfile	CLOSE — False Positive (Docker build secret)
A long regex, UA string, or binary-looking data	CLOSE — False Positive (pattern mismatch)
An actual hardcoded token/key/password string	KEEP OPEN — check liveness if possible
A real URL with embedded credentials (user:pass@host)	KEEP OPEN

For sidekiq-sensitive-url rule: check whether the flagged content is a Redis URL with credentials or just a gem declaration. Almost always a false positive in Gemfile/Gemfile.lock.

Step 3 — If hardcoded secret found:

• Note the exact line and masked value
• Check if secret_liveness column says active — if so, escalate to critical
• Verdict: KEEP OPEN — Hardcoded Secret

---

2b — sast findings

Step 1 — Read the flagged file and lines.

Read CODEBASE_PATH/affected_file at start_line ± 10 lines of context.

Step 2 — Apply rule-specific investigation:

NoSQL injection (NoSQL injection attack possible)

• Trace what the flagged call actually does. Follow the call chain:
  1. What does the flagged function call?
  2. Does it call a NoSQL driver (MongoDB, Redis query, Elasticsearch, Mongoose)?
  3. Or does it call an HTTP client (axios, fetch, HTTParty)?
• Check CODEBASE_PATH/Gemfile and CODEBASE_PATH/package.json for NoSQL drivers.
• If no NoSQL driver exists in the stack: CLOSE — False Positive

SQL injection (SQL injection, string-based query concatenation)

• Read the flagged method. Check:
  1. Is there actual string interpolation into a raw SQL string? ("... #{variable}")
  2. What is the type of the interpolated variable? (Ruby Date, Integer, String?)
  3. Is it executed via select_all, execute, or connection.exec?
  4. Trace the variable back to its source — is it user-controlled?
• If raw interpolation exists but type coercion removes the injection vector: KEEP OPEN — Medium (dangerous pattern, mitigated)
• If raw interpolation with no type coercion on a user-controlled string: KEEP OPEN — High
• If interpolation is of a hardcoded or server-side-only value: CLOSE — False Positive

Unpinned GitHub Actions (3rd party Github Actions should be pinned)

• Read the workflow file at the flagged line.
• Check if uses: has a SHA pin (@abc123def...) or only a tag (@v2, @main).
• Tag-pinned or branch-pinned: KEEP OPEN — Real Finding
• SHA-pinned: CLOSE — False Positive

NODE_AUTH_TOKEN (Use of NODE_AUTH_TOKEN)

• Read the flagged line. If value is ${{ secrets.* }}: CLOSE — False Positive
• If hardcoded token value: KEEP OPEN

Other SAST rules

• Read the flagged lines. Apply judgment: is the finding actually present in the code, or is it a pattern-match on a non-vulnerable construct? Document what you found.

---

2c — open_source / SCA findings

For each SCA finding:

Step 1 — Identify the vulnerable function from the CVE/advisory.

Use your knowledge of the CVE to identify what specific function/class is vulnerable. If you need more detail, use /analyze-cve to do a full dataflow trace.

Step 2 — Check if the package is used in application source code.

Use Grep to search for imports of the package in the application source:

• JS/TS: import .* from 'package-name' or require('package-name')
• Ruby: require 'package' or check Gemfile for direct gem declaration vs transitive

Step 3 — Apply these verdicts:

Scenario	Verdict
Package not imported anywhere in app source (transitive/build-only)	CLOSE — Not Exploitable
Package imported but vulnerable function never called	CLOSE — Not Exploitable
Package imported, vulnerable function called, no user input reaches it	CLOSE — Not Exploitable
Package imported, vulnerable function called with user-controlled input, no sanitization	KEEP OPEN
devDependency only (webpack-dev-server, jest, babel plugins)	CLOSE — Not Exploitable

Step 4 — For complex SCA findings with user-reachable code paths: Run /analyze-cve skill to do a full dataflow trace before rendering the verdict.

---

2d — eol findings

• Read the version file (.ruby-version, check package.json for the framework version).
• Confirm the actual installed version.
• Look up the EOL date from your knowledge.
• If EOL date has passed: KEEP OPEN — Real Finding
• Document the EOL date and recommended upgrade path.

---

Phase 3 — Assign Final Verdicts

For every finding, assign:

Field	Values
recommended_action	KEEP OPEN or CLOSE
close_category	False Positive / File Removed / Not Exploitable / Real Finding
analyst_notes	One sentence — what you found and why
evidence	Specific file:line references and code fragments proving the verdict

---

Phase 4 — Output the Reviewed CSV

Write a new CSV to the same directory as the input CSV, named <original-name>-reviewed.csv.

Column set (lean — do not include empty cloud/VM/container columns from the original):

aikido_issue_id, type, severity, affected_file, related_cve, rule, start_line,
recommended_action, close_category, analyst_notes, evidence

Rules:

• Wrap any field containing commas in double quotes
• Use | (space-pipe-space) as separator within the evidence field — never commas
• Keep analyst_notes to one sentence with no commas
• Every row must have exactly 11 fields

---

Phase 5 — Generate the HTML Evidence Report

Write a self-contained HTML file to the same directory as the CSV, named <project-name>-security-review.html.

HTML structure (follow this exactly)

1. <header>  — project name, analyst, date, branch/source
2. Stats bar — total / close / keep open / false positive / file removed / not exploitable counts
3. Jump nav  — anchor links to each section
4. Full summary table — all findings, colour-coded by severity and action
5. KEEP OPEN section — one detailed card per finding with full code evidence
6. FALSE POSITIVES section — one card per finding with evidence showing why it's a FP
7. FILE REMOVED section — simple table (no code to show)
8. NOT EXPLOITABLE section — one card per finding with code evidence
9. <footer>

Card anatomy (for sections 5, 6, 8)

Each finding card must contain:

• Finding ID, severity badge, type, file:line
• recommended_action badge
• Verdict paragraph — plain English explanation
• Fix (for KEEP OPEN only) — exact command or code change required
• Evidence code block — the actual lines from the file, syntax-highlighted, with line numbers and comments explaining what is or isn't vulnerable

Code block style

Use <pre> blocks with dark background. Add inline <span> highlights:

• class="highlight" — red, for the vulnerable or suspicious line
• class="ok" — green, for the safe/correct pattern
• class="comment" — gray, for explanatory annotations added by the analyst

Severity badge colours

• Critical → red background
• High → orange background
• Medium → yellow background
• Low → green background

Action badge colours

• KEEP OPEN → red
• CLOSE → green

Close category badge colours

• False Positive → blue
• File Removed → purple
• Not Exploitable → green
• Real Finding → red

CSS

Embed all CSS in a <style> block in <head>. No external dependencies — the file must be fully self-contained and openable offline.

Use a clean, professional light theme. Monospace font for file paths and code. System font stack for prose.

---

Phase 6 — Summary to User

After writing both files, output a short summary:

## Triage complete

**CSV:** /path/to/reviewed.csv
**Report:** /path/to/report.html

| Action     | Count |
|------------|-------|
| KEEP OPEN  | N     |
| CLOSE      | N     |
| — False Positive | N |
| — File Removed   | N |
| — Not Exploitable| N |

**Keep open findings:**
- [ID] severity — short description
- ...

---

Rules

• Read every flagged file before rendering a verdict. Never close a finding based on the rule name alone.
• Batch independent reads in parallel — read multiple files in one response when they don't depend on each other.
• Use /analyze-cve for complex SCA findings where user input may reach the vulnerable function. Do not skip this.
• For leaked_secret findings on files that exist: always read the exact flagged lines. Do not assume it's a false positive without looking.
• Preserve line numbers in all code snippets — use the actual line numbers from the source file.
• Do not fabricate code. Only quote lines you have actually read from the file.
• Write the HTML last — after all verdicts are finalized, so the report is complete.
• Do not include empty columns in the output CSV — strip the cloud/VM/container/ARN columns from the original Aikido export.

---

Finding type quick reference

Aikido type	Primary check	Most common outcome
leaked_secret	Does file exist? Then read flagged lines.	False Positive or File Removed
sast — NoSQL	Trace call chain to actual DB driver	Almost always False Positive if MySQL/Postgres stack
sast — SQLi	Read raw SQL method, check interpolation + type coercion	Medium finding, often partially mitigated
sast — Actions	Check for SHA pin vs tag pin	Real finding if tag-pinned
open_source	Grep app source for import, check if vulnerable function used	Often Not Exploitable (transitive/build-only)
eol	Read version file, check EOL date	Always a real finding if EOL date passed