IOC Extraction

Turn messy text into a trustworthy indicator list. The failure modes that matter in a SOC are (1) missing an indicator that was hiding in a defanged or split form, (2) shipping a false-positive indicator that blocks legitimate traffic, and (3) losing the context that tells an analyst why the indicator matters. Everything below is aimed at those three problems.

Workflow

1. Get the raw text. If the input is a file, read it. For PDFs, extract the text first. If the user pasted text directly, use it as-is. Never summarize the source before extracting; extraction works on the full text.
2. Run the extractor script rather than eyeballing regexes. It handles defanged forms (hxxp, [.], (dot), {.}, dot, [@], [://]) and dedupes:

   python scripts/extract_iocs.py <input-file> --format json
   # or from stdin
   cat report.txt | python scripts/extract_iocs.py - --format csv
   # defang for safe sharing in tickets/chat
   python scripts/extract_iocs.py report.txt --format md --defang

   Use --context to include the surrounding sentence for each hit; this is what lets you assign a role (C2, payload host, sender, dropped file) in the next step.
3. Triage the output. The script is deliberately greedy. Now apply judgment:
   • Remove noise: the vendor's own domain, documentation links, example.com, RFC 1918/loopback addresses, Microsoft/Google/CDN infrastructure that appears only as a legitimate reference, version numbers that look like IPs, hashes of known-good files the report mentions for contrast. See references/ioc-types.md for the full noise checklist.
   • Classify each surviving indicator by role (C2, download/staging, phishing sender, phishing landing, exfil destination, dropped file, persistence artifact, scanner source) and confidence (high = the report attributes it directly to the actor; medium = observed but shared/hosting infra; low = mentioned without attribution).
   • Note shelf life: attacker IPs on cloud/VPS rotate in days; hashes are durable but trivially changed; domains sit in between. Say this in the output so the consumer knows what to expire.
4. Produce the deliverable in the format the user's tooling needs (see Output formats). Default to a Markdown table for humans plus a CSV or JSON block for machines.
5. Recommend enrichment and action. For each role, state what to do: block at proxy/DNS, add to EDR watchlist, retro-hunt in SIEM for the last N days, sinkhole, etc. Suggest the enrichment sources from references/enrichment.md, but do not invent enrichment results. If you can actually query a source (via an MCP tool or the user's API), do it and cite it; otherwise mark enrichment as "pending".

Output formats

Always include the source (report title/URL/filename) and extraction date in the output so the list is auditable. Pick the machine format from the user's request:

Markdown table template:

| Indicator (defanged) | Type | Role | Confidence | Notes |
|---|---|---|---|---|
| hxxp://evil[.]example/x.php | url | C2 | high | Cobalt Strike beacon, p.4 |

Things that go wrong

• Split indicators. Reports wrap long URLs across lines and PDF extraction inserts spaces. If the script returns a suspicious fragment (a domain with no TLD, a hash of 63 chars), look at the surrounding text and reassemble by hand.
• Hash type confusion. 32 hex = MD5, 40 = SHA-1, 64 = SHA-256, 128 = SHA-512. A 64-char string can also be an ssdeep-looking blob or a JA3/JA4 fingerprint (JA3 is 32 hex, so it collides with MD5). If the report labels it, trust the label over length.
• Tables of both malicious and benign. Many vendor reports list legitimate tools (PsExec, AnyDesk, rclone) with their hashes. Those are tooling indicators, not malicious files; label them tool and do not recommend blocking without a conversation.
• IPv6, CIDR, and ports. Keep the port with the URL/socket (1.2.3.4:4444) as a note, but the indicator itself is the IP. CIDR ranges are rarely safe to block wholesale; flag them.
• Email addresses as senders vs. victims. A report may include the targeted mailbox. Only sender/reply-to addresses are indicators.
• YARA/Sigma snippets in the report contain strings that look like indicators (mutex names, user agents, paths). Those are detection content, and belong to the detection-engineering skill, but do capture mutexes, named pipes, user agents, and registry keys as host artifacts with type artifact.

Worked example

Input (from a vendor blog post the user pasted):

The loader beacons to hxxps://cdn-sync[.]example/api/v2 and drops C:\Users\Public\svc.exe (SHA256 3a7b...4f1b). Victims received mail from billing[@]invoices-fake[.]com. The malware checks connectivity to www.microsoft.com before running. Analysts can find samples on bazaar.abuse.ch.

Deliverable:

**Source:** "Loader campaign" blog post, 2026-09-17  **Extracted:** 2026-09-17T14:02Z

| Indicator (defanged) | Type | Role | Confidence | Valid until | Notes |
|---|---|---|---|---|---|
| hxxps://cdn-sync[.]example/api/v2 | url | c2 | high | +30d | Beacon endpoint; block full URL, hunt domain |
| cdn-sync[.]example | domain | c2 | high | +90d | Actor-registered per report |
| 3a7b...4f1b | sha256 | dropped-file | high | indefinite | Loader; add to EDR watchlist |
| C:\Users\Public\svc.exe | windows_path | dropped-file | medium | n/a | Hunt only; path is generic |
| billing[@]invoices-fake[.]com | email | phishing-sender | medium | +30d | Mail gateway retro-search 30d |

Dropped as noise: www.microsoft.com (connectivity check), bazaar.abuse.ch (reference).

**Actions**
1. Retro-hunt proxy/DNS for the domain (30d) and EDR for the hash (90d): build the query with `siem-query-authoring`.
2. Mail gateway: search sender 30d; report recipient list to `phishing-analysis` if any delivered.
3. Block URL at proxy; do not sinkhole the domain until passive DNS confirms it is actor-owned.
**Enrichment:** pending (no VT/GreyNoise access in this session).

The point of the example is the shape: every row carries a role, a confidence, and a lifetime, the noise is listed so the reader knows it was considered, and the actions name the skill that does the next step.

Hand-offs

• Hashes and suspicious files: malware-triage for static analysis and YARA.
• CVEs: vulnerability-triage for prioritization; a CVE is not blockable.
• Host artifacts (paths, registry keys, mutexes): detection-engineering for a Sigma rule or threat-hunting for a one-off hunt.
• Retro-search queries across the SIEM: siem-query-authoring.
• The TTPs described around the indicators: mitre-attack-mapping and threat-intel-analysis.

Customization

Teams should edit references/environment.md to list their own allow-listed domains/IP ranges (so they are never emitted as indicators), the exact CSV column order their SIEM lookup expects, and which enrichment sources they have API access to. The script reads an optional references/allowlist.txt (one pattern per line, globs allowed) and drops matches. Add your corporate domains, mail domains, and public ranges there before the first real use; it is the single most effective way to stop your own infrastructure showing up in a blocklist.