redteam-report-template – skill Khác cho agent AI

When to use

Use this skill for client-deliverable reports:

• External red-team engagements with a signed SOW
• Pentest reports going to a CISO / IT-Sec team (not a triager)
• Findings that will be reviewed by both technical and non-technical stakeholders
• Reports that need DOCX/PDF output (not just markdown / platform UI)

Do NOT use for:

• Bug-bounty platform submissions (use report-writing / bugcrowd-reporting instead)
• Quick proof-of-concept memos
• Internal team writeups

---

The 6-section format per finding

This is the canonical structure each finding follows:

## Finding F##: <descriptive title>

**Severity:** Critical / High / Medium / Low / Informational
**Status:** Confirmed / Patched mid-engagement / Suspected (1 signal)
**CVSS 3.1:** <score> (<vector>)
**Affected Asset:** <URL / IP / app name>

### 1. Subject
<One-line statement of the issue. Plain English, no jargon.>

### 2. Observations
<Bulleted list of what was observed during testing. Concrete facts only — no interpretation yet.>
- <Observation 1>
- <Observation 2>
- ...

### 3. Description
<Technical explanation of the vulnerability. 2-4 paragraphs. Reader should understand WHY the observations indicate a vulnerability, what the underlying flaw is.>

### 4. Impact
<What an attacker could achieve. Concrete attacker outcomes, NOT generic CIA triad statements. Tie to the client's business — money, data, reputation, regulatory exposure.>

### 5. Recommendation
<Specific, actionable remediation. Vendor patch, configuration change, code-level fix. Avoid "implement security best practices" — say what specifically.>

### 6. Proof of Concept (PoC)
<Steps to reproduce, numbered. Include the exact HTTP requests, payloads, tools used.>

**Step 1:** <action>

<full HTTP request or curl one-liner>

**Step 2:** <action>

<response excerpt>

**Screenshot:**
Evidence attachment placeholder: screenshots/F<NN>_<descriptive>.png (insert a real local image link only after creating the sanitized file).

---

Before delivery, separate observed impact from plausible but untested impact; repeated exploitation is unnecessary when it would create harm. Protect raw requests, cookies, credentials and identifiers in a restricted evidence appendix, redacting client-facing excerpts. Do not execute PoCs merely to fill this template. A DOCX image-count check is not visual QA: render and inspect pages, redactions, table flow and citations. Cleanup claims require verified artifact inventory, not boilerplate.

Severity & status disciplines

Severity table (client-facing — different from CVSS-only)

| Severity | Business definition | CVSS rough range |

|---|---|---|

| Critical | Direct revenue/data loss without prerequisites | 9.0-10.0 |

| High | Full account/system takeover with limited prerequisites | 7.0-8.9 |

| Medium | Significant data exposure or partial compromise | 4.0-6.9 |

| Low | Information disclosure with limited exploitation path | 0.1-3.9 |

| Informational | Hygiene finding, no immediate exploit | N/A |

Status field (red-team-specific)

This is the field that distinguishes red-team deliverables from bug-bounty reports. Use one of:

• Confirmed — reproduced multiple times, with full PoC
• Confirmed; patched mid-engagement — was reproducible, client patched during the test window (still ship the finding — see mid-engagement-ir-detection)
• Confirmed; partially reproducible — works but needs specific conditions
• Suspected (1 signal) — single indicator, not confirmed (rare — usually drop)
• Out-of-band — finding from passive recon, not actively tested

---

Mistakes to avoid (from authorized-engagement)

1. Don't retract findings that stopped reproducing

If a previously demonstrated finding stops reproducing, preserve the original evidence and timestamps. Record the current retest result; possible causes include a patch, changed state, rate limits, configuration or initial measurement error. Use "patched" only with supporting change/retest evidence, not inference from failure alone. See mid-engagement-ir-detection.

2. Don't hedge in the Impact section

Bad: "An attacker could potentially be able to access user data, which may lead to..."

Good: "An attacker reads any user's profile data. Demonstrated on test user [email protected] at 14:22 IST."

3. Don't generic-CIA the impact

Bad: "Loss of confidentiality and integrity of customer data"

Good: "Read access to 247,000 customer records including PAN cards, addresses, GST numbers. Any legal notification duty needs current jurisdiction-specific counsel review; this technical evidence alone does not establish a statutory deadline."

4. Don't list every recon finding as a finding

Recon notes (subdomains found, ports open, technologies fingerprinted) belong in a separate Recon / Attack Surface appendix, not the findings list. A finding must have an attacker-attainable outcome.

5. Don't bury the PoC

Each finding MUST have reproducible steps. The PoC section is what proves the finding to a skeptical reader. If you can't write the PoC clearly, the finding probably isn't ready to ship.

---

Document-level structure

1. Executive Summary (1 page, non-technical)
   - Engagement overview (dates, scope)
   - Risk posture summary (heat-map: <X critical, Y high, Z medium...>)
   - Top 3 strategic recommendations
   - Comparison to industry baseline (optional)

2. Engagement Details
   - Scope (in-scope, out-of-scope, exclusions)
   - Methodology (recon → exploit → reporting; or align with PTES / OSSTMM)
   - Tools used
   - Timeline (start / end / key milestones)
   - Team

3. Risk Summary Table
   | F# | Title | Severity | Status |
   |----|-------|----------|--------|
   | F01 | ... | Critical | Confirmed |
   ...

4. Findings (one per ## section, in severity order — Critical first)

5. Attack Surface / Recon Appendix
   - Subdomains discovered
   - Open ports / services
   - Technology fingerprints
   - APKs found
   - Credentials in breach corpora (count + sample only — redact)
   - Identity-fabric map (IdP, MFA posture)

6. Indicators of Compromise (IoCs)
   - Source IPs used during testing (so SOC can correlate)
   - User-Agent strings
   - Test accounts created
   - Files uploaded (with cleanup status)

7. Cleanup Statement
   - Confirmation that all test artifacts (accounts, uploads, persistence) were removed
   - Outstanding cleanup items requiring client action

8. Appendices (raw output, screenshots index, full target list)

---

DOCX generation pipeline (markdown → docx with embedded images)

# Prerequisite: pandoc installed; python-docx separately required for the optional check.
# No reference.docx is bundled; provide an actually reviewed local file only if available.
brew install pandoc

# Convert
pandoc REPORT_FINAL.md \
  -o REPORT_FINAL.docx \
  --resource-path=engagement_log/poc \
  --toc \
  --toc-depth=2 \
  --highlight-style=tango

# Verify image count
python3 -c "
from docx import Document
d = Document('REPORT_FINAL.docx')
imgs = [r for r in d.part.rels.values() if 'image' in r.target_ref]
print(f'Embedded images: {len(imgs)}')
print(f'Paragraphs: {len(d.paragraphs)}')
print(f'Headings: {sum(1 for p in d.paragraphs if p.style.name.startswith(\"Heading\"))}')
"

Image filename convention

screenshots/F<NN>_<descriptive>.png

Examples:
F01_locked_accounts.png
F02a_saml_landing.png
F02b_saml_ca_block_page.png
F03_sqli_timing_chart.png
F15_saml_metadata.png

Variants get letter suffixes (F02a, F02b). Always zero-pad finding number.

---

Writing tone — for client deliverables

| Section | Tone |

|---|---|

| Subject | Plain English, jargon-free, 1 line |

| Observations | Bulleted facts, past tense ("observed that...") |

| Description | Technical but accessible; assume CISO reader |

| Impact | Business-translated; tie to revenue/regulation |

| Recommendation | Imperative, specific, actionable |

| PoC | Operator-level technical; copy-pasteable |

Always:

• Use past tense for observations ("The endpoint returned a 200 status code")
• Use present tense for descriptions of the flaw ("The application does not validate...")
• Use imperative for recommendations ("Apply patch ... by ...")
• Number reproduction steps; never "first... then... also..."

Never:

• "Could potentially" — prove it or drop it
• "It might be possible" — same
• "We recommend implementing security best practices" — say which one specifically
• "The application is vulnerable to..." without saying what specifically

---

Audience translation — same finding, different framing

Example: hardcoded JWT in APK

| Section | Technical framing | CISO framing | Board framing |

|---|---|---|---|

| Impact | "JWT signing key extracted from APK enables forging admin tokens" | "Anyone with the customer-facing mobile app can read any customer's invoice" | "A leaked secret in our mobile app lets attackers impersonate us

… (nội dung đầy đủ có trong gói tải về)