Threat model attacked before the design is built

For: Security architect or product security engineer reviewing a new system design

Pattern: Adversarial reviewNeeds live modelsDesigned for 6 to 60 agents

The pain today

Threat models are written by the people who designed the system, so they list the threats the design already handles. The missing ones surface in a penetration test, after the build.

The ask

I attached the design document, the data flow descriptions and our draft threat model for the new customer portal. Have one team defend the mitigations and another try to break them. Show me which threats survive, with the design section that leaves the door open.

Plain words, as you would say it to a colleague. Edit it to fit your case before you send it.

What you attach or connect

  • Design document and data flow descriptions
  • Draft threat model with mitigations
  • Authentication model and trust boundaries
  • Security baseline the design must meet

The unit of work

One worker task per one trust boundary or data flow.

Why a swarm fits

Each boundary can be attacked and defended from its own section of the design. Splitting by boundary keeps context small; opposing teams from different families do not share the designer's blind spots.

Not for

Finding exploitable bugs in running code: that needs testing tools and a tester, not a document review.

The decision tree

5 typed decisions, each with an action for every answer

At fixed moments in a run, the engine puts one narrow question to a decision model. The decision model never writes text: it answers yes or no with a probability, picks from listed options, or gives a score, about a small slice of the material. The engine then does exactly what this tree says, which is what makes the run auditable. The thresholds are the template's design values, not measured results.

  1. Planner, while planning

    Split or keep serialYes or no, with a probability

    While planning

    Does this data flow cross a trust boundary, such as browser to service or service to a third party, so that it deserves its own attack and defence pair?

    Sees only: One data flow description and the list of trust boundaries

    Why: Spends adversarial pairs on boundaries and leaves internal plumbing to one pass.

    • Yes: 0.55 or higherthenSplit the work
    • Unsure: 0.25 up to 0.55thenSplit the work
    • No: below 0.25thenKeep serial
  2. After workers, the judge checks

    Evidence checkYes or no, with a probability

    After a worker answers

    Does the cited design section permit the attacker's step, for example by taking a record identifier from the client with no stated ownership check?

    Sees only: The attack's claimed step and the cited design section

    Why: Attacks resting on facts the documents do not contain become questions, not findings.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.50 up to 0.85thenMark unresolved
    • No: below 0.50thenReject and retry
  3. Evidence checkYes or no, with a probability

    After a worker answers

    Is the mitigation the defence relies on written in the design or the baseline in words that cover this attack path, rather than assumed?

    Sees only: The defence's cited passage and the attack path it answers

    Why: Assumed mitigations are exactly what a threat model written by the designers hides.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.50 up to 0.85thenEscalate to a strong model
    • No: below 0.50thenReject and retry
  4. Reconciler, while merging

    Conflict checkA choice among options

    While reconciling

    Set against the documented mitigation, does the attack still succeed?

    Sees only: The verified attack step and the verified mitigation passage

    Why: Only threats that survive a ruling from a third family reach the architect.

    • The mitigation answers the attackthenAccept
    • The mitigation is partial or covers another paththenEscalate to a strong model
    • It depends on a fact the documents do not statethenMark unresolved
  5. Accountable person, before anything is settled

    Person decidesYes or no, with a probability

    Before anything is reported as settled

    Would this surviving threat expose customer data, credentials or another tenant's resources if the design were built as written?

    Sees only: One surviving threat with its cited design section

    Why: Accepting, mitigating or blocking on a threat is the security architect's decision.

    Accountable: The security architect decides which surviving threats are accepted, mitigated or block the launch.

    • Yes: 0.35 or higherthenAsk a person
    • Unsure: 0.10 up to 0.35thenAsk a person
    • No: below 0.10thenAccept

The fleet: who does what

Model tiers by role, not brands: you choose the models. Strong reasoning models plan and reconcile, small fast models do the wide work, and the judge is a decision model from a different family, so it does not share the workers' blind spots.

  1. Planner

    A strong reasoning model lists trust boundaries and data flows and writes an attack and a defence brief for each.

    Decisions here:1. Split or keep serial

  2. Workers

    Attack and defence workers from two different families argue each boundary, citing only the design text.

    Designed for 6 to 60 agents, one worker task per one trust boundary or data flow. Each worker receives only its own unit.

  3. Judge, from a different model family

    A decision model from a third family rules whether the cited mitigation actually answers the cited attack.

    Decisions here:2. Evidence check3. Evidence check

  4. Reconciler

    A strong reasoning model keeps surviving threats, records refuted ones with the ruling and flags undocumented assumptions.

    Decisions here:4. Conflict check

  5. Accountable person

    The security architect decides which surviving threats are accepted, mitigated or block the launch.

    Decisions here:5. Person decides

Checked before anything is accepted

  • Attacks must cite the design section that permits them
  • Defences must cite a documented mitigation, not an assumed one
  • Threats resting on facts absent from the documents become questions
  • Refuted threats stay in an appendix with the ruling

What comes back

  • Surviving threats by boundary with the cited design section
  • Mitigations claimed but not documented
  • Questions the design does not answer
  • Refuted threats and why

What to measure

  • Surviving threats the architect accepts as real
  • Penetration test findings the review missed
  • Undocumented assumptions found
  • Cost per boundary reviewed

Names of measures only. No result is claimed for this template.

Templates open in the workspace chat with the ask filled in. Nothing runs until you send it.

Get early accessSign in to use

A night of alerts triaged into incidents with evidence

For: Security operations lead or on-call analyst starting the morning shift

Overnight the queue fills with alerts, most of them duplicates or noise.

Pattern: Map, verify, reduceNeeds a connector6 decisionsDesigned for 30 to 1,000 agents

Root-cause hypotheses tested against log windows

For: Site reliability engineer or incident commander writing the review after an outage

After an outage the first plausible story wins.

Pattern: Cross-examinationNeeds scale6 decisionsDesigned for 20 to 500 agents

Access recertification across every account and role

For: IT security manager or identity owner running the periodic access review

Managers receive a spreadsheet of entitlements and approve all of it, because nobody can judge that many rows.

Pattern: Map, verify, reduceNeeds scale5 decisionsDesigned for 40 to 1,000 agents