agentleFS
davidmatousekdavidmatousek/tachi/.claude/skills/tachi-risk-scoring/SKILL.md

tachi-risk-scoring

Skill95 starsChanged 37 days agoApache-2.0View on GitHub

Domain knowledge for quantitative risk scoring — four-dimensional scoring model (CVSS 3.1, exploitability, scalability, reachability), CVSS base vector mappings, composite score formulas, severity band thresholds, and governance field derivation rules. Consumed by the risk-scorer agent during scoring pipeline execution.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 0 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    When a finding has delta_status: UNCHANGED, all scoring fields are inherited verbatim from the baseline run:

    line 29
  2. Very specific:

    When to inherit: Only when delta_status is UNCHANGED.

    line 41
  3. Very specific:

    Every scored finding includes score_source (per schemas/risk-scoring.yaml v1.1):

    line 47
  4. Very specific:

    When scoring NEW findings from Phase 2 isolated discovery, the CVSS base score is constrained within ±1.0 of the category default.

    line 54
  5. Very specific:

    When no baseline is present, all findings receive score_source: "fresh" and are scored using the standard 4-dimensional model.

    line 76
  6. Very specific:

    When scoring findings with delta_status from a baseline-aware run, governance fields follow carry-forward rules to preserve human-assigned …

    line 80

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

tachi-risk-scoring

Quantitative risk scoring domain knowledge extracted from the tachi risk-scorer agent. This skill provides the reference tables, formulas, and assessment criteria that the risk-scorer uses to transform qualitative threat findings into data-backed risk scores.

Domain Overview

The risk scoring model assesses each threat finding on four dimensions:

  1. CVSS 3.1 Base Score (weight: 0.35) -- Inherent vulnerability severity using standard CVSS 3.1 vector analysis, with AI-specific refinements for agentic and LLM threat categories
  2. Exploitability (weight: 0.30) -- Practical attack feasibility across four sub-dimensions: known techniques, attack complexity, tooling availability, and skill level
  3. Scalability (weight: 0.15) -- Blast radius and operational economics across four sub-dimensions: scriptability, target scope, resource requirements, and detection difficulty
  4. Reachability (weight: 0.20) -- Architecture-aware attack surface exposure derived from trust zone position, zone name analysis, and architecture barrier adjustments

Dimensional scores combine into a weighted composite score (0.0-10.0), which maps to a severity band (Critical/High/Medium/Low) that drives governance fields (SLA, disposition, review date).

AIVSS Relationship

Tachi's four-dimensional composite scoring model diverges from OWASP AIVSS at the present time. AIVSS v0.8 (the latest published version, with its public review period opening 2026-04-16) builds on CVSS v4.0 and uses an amplification model in which 10 Agentic AI Risk Factors uplift the CVSS baseline up to a 10.0 ceiling, then a mitigation factor scales the result. Tachi uses CVSS 3.1 and a weighted-sum composite across four operational dimensions (CVSS, Exploitability, Reachability, Scalability). The two formulas produce measurably different scores on the same finding — see ADR-024 for the three-surface comparison (dimensions, formula, severity bands), the worked examples that quantify the divergence, and the re-evaluation trigger tied to AIVSS reaching stable v1.0 with at least one external adopter case study. Severity bands are aligned (Surface C overlap), so downstream consumers (governance, SLA, color coding) behave identically across the two frameworks' outputs.

Baseline-Aware Scoring Rules

Score Inheritance

When a finding has delta_status: UNCHANGED, all scoring fields are inherited verbatim from the baseline run:

FieldTreatmentRationale
cvss_baseCopy verbatimZero drift guarantee
cvss_vectorCopy verbatimAudit consistency
exploitabilityCopy verbatimAssessment unchanged
scalabilityCopy verbatimAssessment unchanged
reachabilityCopy verbatimArchitecture unchanged
composite_scoreCopy verbatimDerived from unchanged inputs
severity_bandCopy verbatimDerived from unchanged composite

When to inherit: Only when delta_status is UNCHANGED. The finding's component, threat, likelihood, and impact must all match the baseline exactly.

When NOT to inherit: UPDATED findings are re-scored fresh using the full 4-dimensional model. NEW findings are scored fresh. RESOLVED findings retain last-known scores but are not actively scored.

Score Source Field

Every scored finding includes score_source (per schemas/risk-scoring.yaml v1.1):

  • "inherited" — Scores copied from baseline (UNCHANGED, RESOLVED)
  • "fresh" — Scores computed this run (NEW, UPDATED)

Bounded Scoring for NEW Findings

When scoring NEW findings from Phase 2 isolated discovery, the CVSS base score is constrained within ±1.0 of the category default. This prevents score drift from LLM non-determinism while still allowing meaningful differentiation between findings in the same category.

Formula: clamped_score = clamp(assessed_score, category_default - 1.0, category_default + 1.0), where clamp(v, min, max) returns min if v < min, max if v > max, and v otherwise. The outer bounds are always [0.0, 10.0].

Applicability:

Delta StatusBounding Applied?Rationale
NEWYesNo established context — bound to category default to prevent outlier scores
UPDATEDNoRe-scored fresh with full context — established finding with history
UNCHANGEDNoScores inherited verbatim — no scoring occurs
RESOLVEDNoScores retained from baseline — no scoring occurs

Edge cases at CVSS extremes:

  • Category default 9.5 → bounded range 8.5–10.0 (CVSS ceiling caps upper bound)
  • Category default 1.0 → bounded range 0.0–2.0 (CVSS floor caps lower bound)
  • Category default 0.5 → bounded range 0.0–1.5 (floor applies)

When no baseline is present: Bounding does not apply. All findings are scored without constraints using the standard 4-dimensional model.

First Run Behavior

When no baseline is present, all findings receive score_source: "fresh" and are scored using the standard 4-dimensional model. No behavioral change from pre-baseline pipeline.

Governance Field Persistence

When scoring findings with delta_status from a baseline-aware run, governance fields follow carry-forward rules to preserve human-assigned values across pipeline runs.

Core principle: risk_owner is a human-assigned field and is never auto-overwritten by the scoring pipeline. SLA and disposition are only recalculated when the severity band changes.

Carry-forward by delta status:

Delta Statusrisk_ownerremediation_slarisk_dispositionreview_date
UNCHANGEDPreserve from baselinePreserve from baselinePreserve from baselinePreserve from baseline
UPDATED (same band)Preserve from baselinePreserve from baselinePreserve from baselinePreserve from baseline
UPDATED (band changed)Preserve from baselineRecalculate per new bandRecalculate per new bandToday + new SLA
NEW"Unassigned"Per severity mappingPer severity mappingToday + SLA
RESOLVEDPreserve from baselinePreserve from baselinePreserve from baselinePreserve from baseline

SLA recalculation trigger: Compare the baseline severity band to the current severity band. If different, recalculate remediation_sla and review_date using the new band's mappings. The risk_disposition threshold is: Critical/High → "Mitigate", Medium/Low → "Review".

Baseline governance detection: Parse the baseline risk-scores.md governance fields by matching on finding ID. If baseline risk-scores.md is unavailable, assign fresh governance fields for all findings (graceful degradation).

Loading Table

Reference files are loaded on-demand by the risk-scorer agent at specific workflow phases using the Read tool (per ADR-002).

ReferenceFileLoad When
Scoring Dimensionsreferences/scoring-dimensions.mdEntering exploitability or scalability assessment phases (Sections 4-5)
CVSS Vectorsreferences/cvss-vectors.mdEntering CVSS 3.1 base scoring phase (Section 3)
Severity Bandsreferences/severity-bands.mdEntering composite calculation or governance field generation phases (Sections 7-8)
Trust Zonesreferences/trust-zones.mdEntering trust zone extraction phase (Section 2)
Reachability Analysisreferences/reachability-analysis.mdEntering reachability assessment phase (Section 6)
Output Formattingreferences/output-formatting.mdEntering markdown output generation phase (Section 9)

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.