Threat Intelligence & CTI

Threat Intelligence & CTI

Purpose

Enable Claude to turn raw observations into finished intelligence — assessments a defender can act on and decision-makers can trust. This skill governs the whole intelligence cycle: framing requirements, collecting and normalizing indicators, applying structured analytic models, scoring sources and confidence, tracking actors and campaigns, and disseminating in machine-readable (STIX/MISP) and human-readable (report) form.

This is distinct from Skill 06 (Threat Hunting): hunting uses intelligence to search an environment for adversary activity; this skill produces and manages the intelligence itself. It is also distinct from Skill 05 (Malware Analysis), which produces the technical facts this skill contextualizes and disseminates.

Activation Triggers

This skill activates when the user asks about:

• Extracting, defanging/refanging, normalizing, or deduplicating IOCs from reports, emails, or feeds
• Producing a STIX 2.1 bundle, a MISP event, or a TAXII-servable indicator set
• Scoring the reliability of a source or the confidence of an assessment (Admiralty/NATO code, words of estimative probability)
• Applying the Diamond Model, Cyber Kill Chain, or MITRE ATT&CK to structure an intrusion
• Tracking or clustering a threat actor / campaign, or reasoning about attribution
• Writing a tactical, operational, or strategic threat intelligence report or an intelligence estimate
• Setting Priority Intelligence Requirements (PIRs) or building a collection plan
• TLP marking, intelligence dissemination, or feed aging/decay and false-positive suppression
• Enriching an indicator (WHOIS/passive DNS/reputation) or pivoting from one indicator to related infrastructure

Prerequisites

python3 --version          # 3.10+; standard library only for the core script
pip install requests       # optional — only for --enrich (live reputation/WHOIS lookups)

Optional enhanced tooling:

• misp / PyMISP — event creation and sharing on a MISP instance
• stix2 / taxii2-client (OASIS) — richer STIX object modeling and TAXII push/pull
• opencti — CTI platform for actor/campaign knowledge-graph management
• A passive-DNS / reputation provider (VirusTotal, Shodan, GreyNoise, urlscan) for enrichment

Handling live malware and indicators: treat sample hashes, live C2 domains, and payload URLs as hostile. Keep them defanged in prose (hxxp://, evil[.]com), never resolve or fetch them from a production host, and mark sharing scope with TLP before dissemination.

Core Capabilities

1. The Intelligence Cycle & Requirements

Anchor every task to where it sits in the cycle, and never skip framing:

1. Direction — establish Priority Intelligence Requirements (PIRs). A good PIR is a decision-relevant question with a consumer and a deadline (e.g., "Which ransomware groups actively target our sector's ERP stack this quarter?"), not "tell me about threats."
2. Collection — map each PIR to specific sources (internal telemetry, OSINT, commercial feeds, ISAC/sharing communities) and note collection gaps explicitly.
3. Processing — normalize, defang, deduplicate, translate, and structure raw data (see §2).
4. Analysis — apply structured models (§3) and estimative language (§4); separate observation from assessment.
5. Dissemination — deliver in the form the consumer can use (§5): STIX for machines, a brief for executives, a detection for the SOC.
6. Feedback — capture whether the product answered the PIR and refine.

State which stage a request touches and what the governing PIR is before producing output.

2. IOC Extraction, Normalization & Enrichment

When the user provides a report, email, or blob and asks for indicators:

1. Extract IPv4/IPv6, domains, URLs, email addresses, file hashes (MD5/SHA1/SHA256), CVE IDs, ASNs, registry keys, mutexes, and Bitcoin/crypto addresses.
2. Refang then re-defang consistently — accept hxxp, [.], (dot), [at], \.; internally canonicalize; always emit defanged in prose and clean values only inside structured/quoted fields.
3. Normalize — lowercase domains, strip URL fragments/default ports where irrelevant, validate hash length per algorithm, drop obvious noise (RFC 1918/loopback/example.com/documentation ranges) unless the user asks to keep them.
4. Deduplicate & type each indicator, and attach context: first/last seen, the report it came from, the kill-chain phase it maps to, and a confidence.
5. Enrich (optional) — WHOIS/registrar, passive DNS, ASN/geo, reputation, and relationships (this domain resolves to that IP, that IP hosts these other domains) to enable pivoting.
6. Use scripts/cti_processor.py for the automatable extraction/defang/normalize/dedup/STIX-export steps; apply human judgment before publishing anything as a confirmed indicator.

3. Structured Analytic Models

Choose the model that fits the question; layer them rather than treating them as alternatives:

• Diamond Model — for a single intrusion event, populate the four vertices (adversary, capability, infrastructure, victim) plus meta-features (timestamp, phase, result, direction, methodology). Pivot across vertices to expand knowledge (from one capability to the infrastructure that delivered it, etc.).
• Cyber Kill Chain (Lockheed Martin) — sequence observed activity across Recon → Weaponization → Delivery → Exploitation → Installation → C2 → Actions on Objectives. Earlier detection is cheaper; note the earliest phase with a detection opportunity.
• MITRE ATT&CK — map each observed behavior to a technique/sub-technique ID and tactic. This is the lingua franca that hands off to Skills 06, 12, 15, and 22.
• Analysis of Competing Hypotheses (ACH) — when attribution or intent is contested, enumerate hypotheses, list evidence, and score evidence by how well it disconfirms each hypothesis (diagnostic evidence), not how well it confirms your favorite.

4. Source & Confidence Scoring

Separate how much you trust the source from how much you trust the claim, and make both explicit:

• Admiralty / NATO source-reliability code — reliability A–F (Reliable → Cannot be judged) × information credibility 1–6 (Confirmed → Cannot be judged). Report as e.g. B2.
• Words of estimative probability — use calibrated language ("almost certainly", "likely", "roughly even chance", "unlikely") and, where useful, an explicit percentage band; never let "could" masquerade as "will".
• Confidence level (low/medium/high) on each assessment, with the drivers named (source count, corroboration, analytic assumptions).
• Flag assumptions and information gaps in-line — an unstated assumption is the most common failure mode in CTI.

5. Actor & Campaign Tracking + Attribution Discipline

1. Cluster by TTP first, name later. Group activity into an unnamed activity cluster (UNC-style) from shared infrastructure, tooling, and behavior before reaching for a public actor name.
2. Attribution is a spectrum, not a verdict. Distinguish technical attribution (this cluster did this) from actor attribution (this named group) from nation/sponsor attribution (the hardest and most consequential). State the confidence and the evidence class for each level.
3. Beware overlap and false flags — shared tooling, commodity malware, and deliberate false flags make single-indicator attribution unreliable. Weight durable behavioral TTPs over swappable infrastructure.
4. Track campaigns with a consistent internal schema (cluster ID, aliases, associated malware, targeted sectors/geos, first/last activity) and map aliases across vendor naming (MISP galaxies help).

6. Production, Dissemination & Feed Hygiene

• Match the product to the consumer — tactical (IOCs/detections for the SOC), operational (TTPs/campaigns for defenders and IR), strategic (trends/risk for leadership). One input often yields three different products.
• TLP 2.0 marking on every product (TLP:RED / AMBER+STRICT / AMBER / GREEN / CLEAR) — the marking is the sharing contract; state it before distributing.
• Machine-readable dissemination — emit STIX 2.1 bundles (indicators, relationships, marking-definitions) for automated ingestion; MISP events for community sharing; TAXII for pull/push.
• Feed hygiene — indicators decay. Apply aging/scoring so stale indicators expire, suppress known-false-positive infrastructure (CDNs, sinkholes, shared hosting) with an allowlist, and never push an un-triaged feed straight to blocking.

Output Template

# Threat Intelligence Product
**Title:** [Short descriptive title]
**Date:** [YYYY-MM-DD]  |  **TLP:** [RED / AMBER+STRICT / AMBER / GREEN / CLEAR]
**Product type:** [Tactical / Operational / Strategic]
**PIR addressed:** [the requirement this answers]

---

## Bottom Line Up Front (BLUF)
[2–3 sentences of assessment, with an estimative-language judgment and an overall confidence level.]

## Key Assessments
- [Assessment 1] — *Confidence: [Low/Med/High]* — *Source: [Admiralty e.g. B2]*
- [Assessment 2] — ...

## Diamond / Kill Chain Summary
| Vertex / Phase | Observed |
|----------------|----------|
| Adversary | [cluster/actor + attribution confidence] |
| Capability | [malware/tooling + ATT&CK techniques] |
| Infrastructure | [C2/staging + defanged] |
| Victim | [sector/geo/asset] |
| Earliest detection opportunity | [kill-chain phase] |

## Indicators of Compromise
| Indicator (defanged) | Type | Kill-chain phase | Confidence | First seen | Source |
|----------------------|------|------------------|------------|-----------|--------|

## ATT&CK Techniques Observed
| Tactic | Technique (ID) | Evidence |
|--------|----------------|----------|

## Assumptions & Intelligence Gaps
- [Assumption / gap and how it affects confidence]

## Recommended Actions
- [Detection to deploy → Skill 12 / 15] · [Hunt to run → Skill 06] · [Emulation to validate → Skill 22]

Script Reference

cti_processor.py

# Extract, defang, normalize, dedup IOCs from a report and print a summary
python scripts/cti_processor.py --input report.txt

# Emit a STIX 2.1 bundle and a MISP-style event
python scripts/cti_processor.py --input report.txt --stix iocs_stix.json --misp event.json

# Tag every indicator with a TLP marking and an Admiralty source score
python scripts/cti_processor.py --input report.txt --tlp AMBER --source-score B2 --output iocs.json

# Optional live enrichment (adds network calls; requires `requests`)
python scripts/cti_processor.py --input report.txt --enrich --output iocs.json

# Self-contained demonstration on a bundled sample report
python scripts/cti_processor.py --demo

Extracts IPv4/IPv6, domains, URLs, emails, MD5/SHA1/SHA256 hashes, CVEs, ASNs, and crypto addresses; refangs mixed defang styles, canonicalizes and validates, drops private/documentation noise, deduplicates, and exports STIX 2.1 / MISP-style JSON with TLP and Admiralty scoring. Extraction and confidence are heuristic — verify before publishing an indicator as confirmed.

Skill Integration

References

• STIX 2.1 (OASIS) · TAXII 2.1 (OASIS)
• MITRE ATT&CK
• The Diamond Model of Intrusion Analysis (Caltagirone, Pendergast, Betz)
• Lockheed Martin Cyber Kill Chain
• MISP — Open Source Threat Intelligence Platform
• FIRST Traffic Light Protocol (TLP) 2.0
• Admiralty / NATO source-reliability grading
• CIA — Words of Estimative Probability (Sherman Kent)
• Psychology of Intelligence Analysis / ACH (Richards Heuer)