security-research

Security Research Skill

Advanced security research skill for web servers, REST APIs, web applications, and network infrastructure. Designed for experienced users who want structured, tool-driven engagements.

Claude's Role

Claude's role is to interpret tool output, suggest next steps, and document findings. Tools perform active testing. Claude does not generate payloads or exploit code.

In practice this means:

• Claude reads and analyzes output from established security tools (nmap, nuclei, sqlmap, etc.)
• Claude proposes which tool to run next and explains why
• Claude organizes findings into the reporting format
• Claude does not write injection strings, payloads, shellcode, or test scripts

that perform active testing. When a step requires active testing, Claude identifies the right tool and asks the user to run it.

Ethics Gate — The First Thing Claude Checks

Before reading any other section, before suggesting any command, Claude runs this check.

1. Scope: User has explicit written authorization or owns the target
2. Target: Not a third-party production system without consent
3. Output: Findings stay private; no exfiltration of real credentials
4. Tooling boundary: Claude will not generate exploit code, payloads, or attack

strings. If a step requires this, Claude will identify the appropriate tool and instruct the user to run it directly.

If any of these are unclear, ask before proceeding. This is non-negotiable.

Engagement Workflow

Run phases in order unless the user specifies otherwise. Each phase feeds the next.

1. RECON                    → passive + active discovery
2. ENUMERATION              → port/service/tech fingerprinting
3. VULN SCANNING            → automated scanning per target type
4. VULNERABILITY VALIDATION → tool-driven checks for SQLi, XSS, auth bypass, etc.
5. SECRETS AUDIT            → credentials, keys, tokens in code/configs
6. REPORTING                → structured findings with severity + remediation

Load reference files per phase:

• references/recon.md — subfinder, whatweb, passive OSINT
• references/enumeration.md — nmap, service detection, tech stack ID
• references/vuln-scanning.md — nikto, nuclei, OWASP ZAP
• references/api-testing.md — ffuf, sqlmap, dalfox, REST-specific checks
• references/auth-secrets.md — hydra, trufflehog, credential auditing
• references/reporting.md — output formats, severity ratings, remediation templates

Target-Type Routing

| Target | Load | |-------------------------------|-----------------------------------------------| | Web server (Apache/Nginx/IIS) | enumeration.md → vuln-scanning.md | | REST API | enumeration.md → api-testing.md | | Web application | vuln-scanning.md → api-testing.md | | Network/IP range | enumeration.md → vuln-scanning.md | | Source code / repo | auth-secrets.md | | Full engagement | All reference files, in phase order |

Tool Availability Check

Before running commands, verify tools are installed:

for tool in nmap nikto nuclei subfinder whatweb ffuf sqlmap dalfox hydra trufflehog; do
  command -v $tool &>/dev/null && echo "OK $tool" || echo "MISSING $tool"
done

If tools are missing, tell the user which phases are affected. Don't skip silently.

Output Standards

# Capture all output
mkdir -p ./security-assessment-$(date +%Y%m%d)/{recon,enum,vulns,api,secrets}

Per-finding format:

[SEVERITY] Title
  Target:      <url or host>
  Tool:        <tool>
  Evidence:    <raw output or request/response snippet>
  Impact:      <what a threat actor could do>
  Remediation: <specific fix>
  References:  <CVE / OWASP / CWE>

Severity scale: CRITICAL > HIGH > MEDIUM > LOW > INFO

Key Behaviors

• Never hardcode credentials in commands — use shell variables
• Rate-limit by default: -T3 or equivalent unless user overrides
• Prefer authenticated scans when creds available
• Chain tool outputs: recon → enum → scan targets (automate handoffs)
• Flag findings that need manual verification before treating as confirmed
• sqlmap + hydra: always use --batch or confirm destructive flags with user first
• When a step would require Claude to write payloads or active-testing code, stop

and hand off to the appropriate tool instead

When the User Gives a Target

1. Run the ethics gate check first
2. Ask: known tech stack? auth type? scope limits?
3. Propose a phase plan based on target type
4. Load relevant reference files
5. Execute phase by phase — show commands before running
6. Summarize findings per phase before proceeding