Imported from hackersifu/reverse-engineering-skills (.agents/skills/re-ioc-extraction/SKILL.md). Install upstream with npx skills add hackersifu/reverse-engineering-skills --skill re-ioc-extraction. Copyright stays with the author.

re-ioc-extraction

Purpose

Produce a complete, defensive IOC set from analyst-provided evidence and present it in:

1. a Markdown IOC table, and
2. a structured YAML list,
   with traceable evidence only (no guessing, no enrichment, no validation).

Inputs expected

Provide at least one of the following (more is better):

• strings output (ASCII and/or UTF-16)
• sandbox / EDR / detonation logs (process, file, registry, network)
• packet/proxy/DNS logs or extracted HTTP requests
• reverse engineering notes (imports, APIs, decompiled snippets, paths, config blobs)
• known hashes from a trusted source

If evidence is missing, the skill must still produce output for what is present and clearly note gaps.

Optional evidence generation (static only, no execution)

This skill must not execute the sample. If the user requests execution (or execution is required to obtain evidence), pause and notify the engineer to decide whether to proceed in a controlled sandbox.

If the user asks for help generating evidence locally (or you are explicitly allowed to run commands), prefer small, single-purpose commands that are hard to break and easy to copy/paste.

Command hygiene (important)

• Prefer one pattern per command over long one-liners with many quoted arguments.
• Avoid complex nested quotes. If a command needs heavy quoting, break it into multiple commands.
• If a combined command fails (e.g., quoting/EOF errors), rerun as several simple commands and proceed.

Tool availability check (recommended)

command -v file strings sha256sum sha1sum md5sum capa jq || true

file (type/format context)

file "<sample>"

hashes (for reporting / dedupe)

sha256sum "<sample>"
sha1sum "<sample>"
md5sum "<sample>"

strings (baseline)

strings -a "<sample>" | head -n 2000

strings (UTF-16 / wide strings on Windows malware)

strings -el "<sample>" | head -n 2000

Searching extracted strings safely (recommended pattern)

First save strings to a file, then search with one pattern per command:

strings -a "<sample>" > /tmp/sample.strings.txt
strings -el "<sample>" > /tmp/sample.strings.utf16.txt

rg -n -F "http://" /tmp/sample.strings.txt
rg -n -F "https://" /tmp/sample.strings.txt
rg -n -F ".onion" /tmp/sample.strings.txt
rg -n -F "HKEY_" /tmp/sample.strings.utf16.txt
rg -n -F "CurrentVersion\Run" /tmp/sample.strings.utf16.txt

capa (use when available; degrade gracefully)

Use capa only as supporting static context (capabilities, packer/obfuscation hints). It does not directly produce IOCs.

• If capa fails due to rule/signature mismatches, note it and continue with the evidence you do have.
• If you have to pin rules, do so explicitly and record the rules version used in notes.

Non-negotiable rules

1. No hallucinations: only output indicators explicitly present in the evidence.
2. If incomplete/ambiguous (e.g., hxxp://, partial domains, truncated hashes), label candidate / incomplete and state what’s missing.
3. No live validation: do not resolve domains, visit URLs, or test infrastructure. Extraction and normalization only.
4. Every IOC must be traceable: include a verbatim evidence snippet (exact line or tight excerpt) and the source artifact (strings/log/notes).
5. Stay on scope: do not inspect unrelated environment files (shell dotfiles, configs) unless explicitly requested. Focus on the sample and provided evidence.
6. Do not “upgrade” indicators: do not infer C2/persistence/stealer behavior into network IOCs unless the indicator values are present. Behavioral assessment is optional notes only.
7. No execution: do not run the sample or trigger detonation. If execution is requested or required, pause and notify the engineer to decide (and require a controlled sandbox/VM).

Execution required — pause template

If execution is requested/required, respond with:

PAUSE: This step would require executing an unknown sample. I will not proceed automatically.
Please confirm (engineer decision) whether to run in a controlled sandbox/VM, and provide the sandbox constraints (networking, snapshotting, logging tools).
If approved, paste sandbox outputs here and I will extract IOCs from the resulting evidence.

IOC types to extract

Extract only what is present in evidence. Common types include:

• Hashes: MD5 / SHA1 / SHA256
• Network: domains/subdomains, IPs (v4/v6), URLs, URI paths, ports, SNI/Host headers if present
• Email addresses (if present)
• File system: dropped file names, file paths
• Windows persistence: registry keys/values, services, scheduled task names/paths (only if values are present)
• Mutex names
• User-Agent strings
• Process names / command lines (only if explicitly present)
• Certificates/keys: subjects, thumbprints, public keys (only if explicitly present)

Normalization rules

• Domains: lowercase, strip surrounding punctuation, keep subdomains
• URLs: preserve as-seen; if obfuscated (hxxp), include both obfuscated and normalized forms
• IPs: normalize formatting; preserve IPv6
• Hashes: lowercase; do not “fix” length or guess missing characters
• Paths/registry: preserve exactly as provided; do not guess missing segments
• De-obfuscation: only apply mechanical transforms that are explicitly reversible and obvious (e.g., hxxp→http). Do not decode custom encodings unless the decoded value is shown in evidence.

Confidence labels

Use a consistent, limited vocabulary:

• confirmed: directly observed in authoritative output (hash tool output, explicit log lines)
• high / medium / low: observed in evidence but context quality varies (e.g., noisy logs, partial correlation)
• contextual: value is present but not inherently malicious (e.g., PDB path, product name, benign-looking file name)
• candidate: looks like an IOC but may be incomplete/obfuscated/ambiguous
• incomplete: truncated or partial value (state what is missing)

Output (always produce both)

A) IOC table (Markdown)

Include columns:

• Type
• Indicator
• Confidence
• Context (short)
• Evidence (verbatim snippet + where it came from)

Rules:

• One indicator per row.
• Evidence must be an exact line or tight excerpt containing the indicator (do not paraphrase).
• If the same IOC appears in multiple sources, include the best/most authoritative snippet, and optionally list a second snippet in the context.

B) Structured IOC list (YAML)

Schema (strict)

Top-level keys (include only those that have entries):

• hashes
• network
• file_paths
• file_names
• process_names
• registry
• mutexes
• user_agents
• emails
• certificates
• notes (optional)
• static_risk_notes (optional; see below)

Each list entry must include:

• value (string)
• confidence (one of: confirmed, high, medium, low, contextual, candidate, incomplete)
• source (short label such as: "hash command output", "strings output", "UTF-16 strings output", "sandbox log", "network log", "RE notes")
• evidence_snippet (verbatim line/excerpt containing the value)

Additional required fields by type:

• hashes[]: algorithm (md5|sha1|sha256)
• network[]: kind (domain|ipv4|ipv6|url|uri_path|port|sni|host_header)
  • If kind: port, value must be a string like "tcp/443" or "udp/53" when protocol is known; otherwise "port/443".
• registry[]: kind (key|value|data) when known
• certificates[]: kind (subject|thumbprint|public_key) when known

Example (shape only):

hashes:
  - value: "..."
    algorithm: "sha256"
    confidence: "confirmed"
    source: "hash command output"
    evidence_snippet: "sha256sum sample -> ..."

network:
  - kind: "domain"
    value: "example.com"
    confidence: "high"
    source: "sandbox log"
    evidence_snippet: "DNS query: example.com"

Optional: static risk notes (do not invent IOCs)

If the user asks “is it malicious?” you may include a short static_risk_notes section in YAML (and/or a short paragraph after the IOC table) that:

• cites the exact static evidence used (e.g., packing/obfuscation signs, capa capability names, suspicious API surface),
• avoids definitive attribution,
• never manufactures missing IOCs.

Keep it brief and operational (what defenders should treat as risky and why).

Minimal clarification questions (only if required; max 1–3)

Ask only if needed to avoid incorrect extraction:

• “Do you want only confirmed IOCs, or include candidate/contextual values too?”
• “Is this evidence from strings only or do you have sandbox/network logs as well?”
• “Do you want output optimized for a specific consumer (SIEM rule, EDR blocklist, report)?”

Happy path example (regression anchor)

Example evidence (input)

sha256sum sample.exe -> 1111111111111111111111111111111111111111111111111111111111111111
sha1sum sample.exe   -> 2222222222222222222222222222222222222222
md5sum sample.exe    -> 33333333333333333333333333333333

strings:
... hxxp://BadDomain[.]com/path/index.php
... User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64)
... D:\Build\proj\bin\Release\sample.pdb

Expected IOC table (output)

Expected YAML (output)

hashes:
  - value: "1111111111111111111111111111111111111111111111111111111111111111"
    algorithm: "sha256"
    confidence: "confirmed"
    source: "hash command output"
    evidence_snippet: "sha256sum sample.exe -> 1111111111111111111111111111111111111111111111111111111111111111"
  - value: "2222222222222222222222222222222222222222"
    algorithm: "sha1"
    confidence: "confirmed"
    source: "hash command output"
    evidence_snippet: "sha1sum sample.exe   -> 2222222222222222222222222222222222222222"
  - value: "33333333333333333333333333333333"
    algorithm: "md5"
    confidence: "confirmed"
    source: "hash command output"
    evidence_snippet: "md5sum sample.exe    -> 33333333333333333333333333333333"

network:
  - kind: "url"
    value: "hxxp://BadDomain[.]com/path/index.php"
    confidence: "candidate"
    source: "strings output"
    evidence_snippet: "... hxxp://BadDomain[.]com/path/index.php"
  - kind: "url"
    value: "http://baddomain.com/path/index.php"
    confidence: "candidate"
    source: "strings output"
    evidence_snippet: "... hxxp://BadDomain[.]com/path/index.php"

user_agents:
  - value: "Mozilla/5.0 (Windows NT 10.0; Win64; x64)"
    confidence: "contextual"
    source: "strings output"
    evidence_snippet: "... User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64)"

file_paths:
  - value: "D:\\Build\\proj\\bin\\Release\\sample.pdb"
    confidence: "contextual"
    source: "strings output"
    evidence_snippet: "... D:\Build\proj\bin\Release\sample.pdb"

Definition of done

• Every indicator in the table is present in the evidence and has a verbatim snippet.
• YAML conforms to the schema and uses only allowed confidence labels.
• Normalization is applied without guessing missing values.
• No live validation/enrichment was performed.
• If the user asked about maliciousness, any risk notes are brief, evidence-cited, and do not invent IOCs.