Writing intelligence reports

Start from intelligence requirements, write key judgements with calibrated estimative language (ICD 203 and the UK probability yardstick), grade sources with the Admiralty code, structure a report that a CISO and a detection engineer can both act on, and share it safely with TLP 2.0, MISP and STIX 2.1.

FREEBy ·Lesson·13 min read·Intermediate → Advanced
intelligence reportestimative languageAdmiralty codeTLP 2.0STIX 2.1MISP

01Start from intelligence requirements

✓What you will learn
  • Anchor reports to intelligence requirements and the audience that will act on them
  • Write judgements with estimative language and a separate confidence level
  • Grade sources and information with the Admiralty code
  • Structure a report around key judgements, evidence and actions
  • Share safely with TLP 2.0, MISP and STIX 2.1

Intelligence is information analysed to answer a decision-maker's question. Without the question it is just information. Intelligence requirements are those questions, agreed with stakeholders; the most important are priority intelligence requirements (PIRs), for example "Which ransomware groups are targeting UK logistics firms, and how do they get initial access?". Every report should say which requirement it answers.

LevelAudienceQuestionFormat
StrategicBoard, CISOWhat threats matter to the business, and what should we invest in?One page, plain language, trends and risk
OperationalIR lead, SOC managerWho is coming, how, and when? What campaign is this?Campaign report with actor, TTPs and timeline
TacticalSOC analysts, detection engineersWhat do we hunt, detect and block today?IOCs, ATT&CK mappings, queries and rules

The intelligence cycle keeps this honest: direction (requirements), collection, processing, analysis, dissemination, and feedback. Feedback is the step most teams skip, and it is how you learn that the board wanted a risk statement, not a list of 400 hashes.

02Confidence and estimative language

Analysts rarely know for certain. Estimative language expresses how likely you judge something to be, using words with agreed probability ranges so that "likely" means the same thing to the writer and the reader. The idea goes back to Sherman Kent at the CIA, who found that readers of one estimate took "serious possibility" to mean anything from about 20% to 80%.

US ICD 203 termProbabilityUK PHIA yardstick termProbability
Almost no chance1-5%Remote0-5%
Very unlikely5-20%Highly unlikely10-20%
Unlikely20-45%Unlikely25-35%
Roughly even chance45-55%Realistic possibility40-50%
Likely55-80%Likely or probably55-75%
Very likely80-95%Highly likely80-90%
Almost certain95-99%Almost certain95-100%

Likelihood and confidence are different things. Likelihood is your estimate of the event. Confidence (low, moderate, high) is how much you trust the basis for that estimate: source quality, corroboration, gaps. "We assess it is likely that the actor will target our VPN again, with moderate confidence" is a complete judgement. The UK yardstick deliberately leaves gaps between ranges so that writers commit to one term.

Admiralty code: source reliabilityInformation credibility
A: completely reliable1: confirmed by other sources
B: usually reliable2: probably true
C: fairly reliable3: possibly true
D: not usually reliable4: doubtful
E: unreliable5: improbable
F: reliability cannot be judged6: truth cannot be judged

The Admiralty code (also called the NATO system) grades each piece of information on two independent axes: how reliable the source has been historically, and how credible this item is. A long-trusted vendor passing on an uncorroborated forum rumour is B4; a first-time anonymous tip that matches your own logs is F1. Keeping the axes separate stops a good source's reputation from laundering a weak claim.

iWords to avoid

"Could", "may" and "might" commit to nothing: almost anything could happen. "Possibly" and "perhaps" are not on either scale. If you cannot attach a calibrated term, say what evidence would let you, and put it under intelligence gaps.

03Structure of a good report

Readers decide in the first paragraph whether to keep reading. Lead with BLUF (bottom line up front): the key judgements and what to do, then the evidence. Technical appendices carry the detail for the people who need it.

A one-page report skeletontext
TLP:AMBER  CYART-CTI-2026-041   2026-11-14   Answers PIR-3 (credential phishing against staff SSO)

KEY JUDGEMENTS
1. A financially motivated actor is running a credential-phishing campaign impersonating our Okta SSO.
   We assess it is almost certain the campaign targets us specifically (brand-specific domains), high confidence.
2. It is likely (55-80%) that more staff will receive lures this week: four lookalike domains are registered
   and certified but unused. Moderate confidence: based on CT and passive DNS, no access to actor tooling.
3. We have no evidence of successful account takeover as of 2026-11-14 09:00 UTC (intelligence gap:
   personal-device logins are not logged).

RECOMMENDED ACTIONS
- Block the 6 domains and 2 IPs in Appendix A at mail, proxy and DNS (expiry 30 days).
- Enforce phishing-resistant MFA (FIDO2) for the finance and HR groups first.
- SOC: hunt sign-ins from AS64501 and new MFA device registrations since 2026-11-09 (queries in Appendix B).

DETAILS        timeline, infrastructure analysis, ATT&CK mapping (T1566.002, T1078, T1556.006)
SOURCES        internal mail logs (A1), passive DNS provider (B2), crt.sh (A1), vendor report VR-2291 (B3)
GAPS           no visibility of personal devices; actor attribution not attempted
APPENDIX A     defanged IOCs with first-seen and expiry       APPENDIX B   KQL and SPL hunts
▭On the whiteboard · What each part is for6 points
01
header
TLP, unique ID, date, the requirement it answers
02
key judgements
Two to five numbered statements with likelihood and confidence
03
actions
Specific, owned, time-bound: who blocks what by when
04
details
Evidence, timeline, ATT&CK mapping for the analysts who will check you
05
sources and gaps
Admiralty grades, what you could not see, what would change your mind
06
appendices
Machine-readable IOCs and hunts, kept out of the reader's way
!Common report failures

Burying the judgement under five pages of timeline. Using "could" and "may" everywhere. Attributing to a named nation-state on weak, single-source evidence. Pasting 300 IOCs with no context or expiry. And writing for yourself: a CISO needs the risk and the decision, not the JARM hash.

04Sharing: TLP 2.0, MISP and STIX

The Traffic Light Protocol (TLP) tells recipients how far they may share what you send. FIRST published TLP 2.0 in August 2022; it renamed WHITE to CLEAR and added AMBER+STRICT. The label goes at the top of the report and on every exported indicator.

LabelRecipients may share withTypical use
TLP:REDNo one: for the named recipients only, often in a meetingInformation that could compromise an ongoing investigation
TLP:AMBER+STRICTPeople inside their own organisation only, on a need-to-know basisVictim-specific details shared with a partner
TLP:AMBERTheir organisation and its clients, on a need-to-know basisCampaign details for peers who need to protect themselves
TLP:GREENTheir community (for example an ISAC), but not publiclyIndicators and TTPs for sector peers
TLP:CLEARAnyone, without restrictionPublic advisories and blog posts

MISP is the most widely used open-source threat sharing platform. An event groups attributes (typed indicators such as domain, ip-dst, sha256, each with a to_ids flag saying whether it is fit for automated detection), objects (structured groups of attributes), tags (tlp:amber) and galaxies (ATT&CK techniques, actor clusters). Distribution levels control sharing: your organisation only, this community, connected communities, all communities, or a named sharing group.

STIX 2.1 is the OASIS standard data format for intelligence, and TAXII 2.1 is the HTTPS protocol for exchanging it. STIX models domain objects (indicator, malware, intrusion-set, attack-pattern, report), cyber observables (domain-name, ipv4-addr, file) and relationships between them. TLP is carried as marking definitions referenced from each object.

STIX 2.1: an indicator with a validity window and an indicates relationshipjson
{
  "type": "bundle",
  "id": "bundle--6f1e2c3a-8b4d-4e5f-9a0b-1c2d3e4f5a6b",
  "objects": [
    {
      "type": "indicator",
      "spec_version": "2.1",
      "id": "indicator--0c9b8a7d-6e5f-4a3b-8c2d-1e0f9a8b7c6d",
      "created": "2026-11-14T09:30:00.000Z",
      "modified": "2026-11-14T09:30:00.000Z",
      "name": "Okta SSO phishing domain",
      "indicator_types": ["malicious-activity"],
      "pattern": "[domain-name:value = 'cyart-sso-login.com']",
      "pattern_type": "stix",
      "valid_from": "2026-11-09T13:51:22Z",
      "valid_until": "2026-12-14T00:00:00Z",
      "confidence": 85
    },
    {
      "type": "relationship",
      "spec_version": "2.1",
      "id": "relationship--4d3c2b1a-0f9e-4d8c-b7a6-5f4e3d2c1b0a",
      "created": "2026-11-14T09:30:00.000Z",
      "modified": "2026-11-14T09:30:00.000Z",
      "relationship_type": "indicates",
      "source_ref": "indicator--0c9b8a7d-6e5f-4a3b-8c2d-1e0f9a8b7c6d",
      "target_ref": "attack-pattern--7e3a1c52-9b4d-4f6e-a8c0-2d1b3f5e7a90"
    }
  ]
}

Note valid_until: the format has expiry built in, which is the indicator half-life from the first lesson in machine-readable form. Set it. The relationship's target_ref points at an attack-pattern object (the ATT&CK technique, here the phishing technique) defined elsewhere in your bundle or collection.

“How to say it in the room

"Every report I write answers a named requirement, leads with numbered key judgements that carry both a calibrated likelihood and a confidence level, grades sources with the Admiralty code, and ends with owned actions. Indicators go out as STIX or MISP with a TLP label and an expiry, never as a pasted list."

◎Interview radar

CTI interviews test whether you can separate likelihood from confidence, name the TLP 2.0 labels (including CLEAR and AMBER+STRICT), and explain the Admiralty code's two axes. Writing exercises are common: you get raw notes and 30 minutes to produce key judgements. Lead with the bottom line.

Test yourself

A report says: "We assess it is likely the actor will return, with low confidence." What does this communicate?

Q1AMBER or AMBER+STRICT for a partner?+

AMBER+STRICT if the content could identify a victim or client and must stay inside the recipient organisation; AMBER if the recipient may need to pass it to clients they protect, such as an MSSP.

Q2When do you set to_ids on a MISP attribute?+

Only when the value is specific enough to detect on without false positives: a phishing domain yes, a shared cloud IP or a common file name no. Context-only attributes keep to_ids off.

Q3How long should a report be?+

As long as the audience needs and no longer. Key judgements and actions fit on one page for any audience; details and appendices can be as long as the evidence requires, because only the people checking your work read them.

Finished this one?
Tick it and it shows as done on the roadmap and the course.