Skip to main content
subagent CX Codex

risk-scorer — Claude Code subagent: what it does, how to install

risk-scorer scores vulnerabilities using CVSS, EPSS, and CISA KEV to build a justified remediation queue aligned with business context.

MIT ★ 2,482 updated 2026-10-06

Install into Claude Code · Manually: put the files in place All methods and platforms ↓
.claude/agents/risk-scorer.md
Open on GitHub

What risk-scorer does

This subagent processes security findings by calculating detailed CVSS v3.1 or v4.0 vector strings. It enriches the base score with real-world threat indicators, including the 30-day exploitation probability from EPSS and active weaponization status from the CISA KEV catalog. Because of real-world activity, a medium CVSS finding listed in KEV frequently outranks a theoretical high-severity bug.

The final rank accounts for exposure, asset criticality, and compensating security controls. Each finding receives a priority tier from P1 to P4 with documented rationale covering the attacker, defender, and business perspectives. If the findings.sh helper utility is discovered on your system, the agent queries vulnerabilities via findings.sh list vulns and logs the resulting vector and score back automatically.

The scope is limited strictly to scoring and ranking: finding discovery belongs to test agents, multi-step sequencing goes to attack-planner, and report drafting delegates to report-generator.

What it helps with

  • build justified CVSS v3.1 and v4.0 vector strings for identified findings
  • enrich vulnerability records with EPSS probabilities and CISA KEV data
  • adjust remediation ordering according to exposure and compensating controls
  • assign clear P1 through P4 priority tiers backed by attacker and defender rationale

Example prompts for the agent

  • Score the current vulnerability findings using CVSS v4.0 and check against CISA KEV
  • Prioritize our web application flaws considering internet exposure and internal controls
  • Fetch findings using findings.sh, assign P1-P4 tiers, and log the results back

How to install risk-scorer

Manually: put the files in place

.claude/agents/risk-scorer.md

For all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention

Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean

Questions about risk-scorer

How to install risk-scorer in Claude Code?

Place the subagent definition file into your project at .claude/agents/risk-scorer.md.

What does risk-scorer do?

It turns security findings into prioritized P1–P4 remediation lists by building CVSS vectors, adding EPSS and KEV context, and factoring in asset criticality.

Does risk-scorer discover vulnerabilities directly?

No, it focuses solely on scoring and ranking existing findings. Discovery, attack-chain sequencing, and report creation belong to separate testing and reporting agents.

Source text

risk-scorer.md · MIT
---
name: risk-scorer
description: Delegates to this agent when the user wants to score and prioritize findings — build CVSS 3.1/4.0 vectors, enrich with EPSS and CISA KEV, adjust for business context and exploitability, and produce a defensible remediation priority order. Distinct from attack-planner (attack-path sequencing) and report-generator (report assembly).
tools:
  - Read
  - Write
  - Edit
  - Grep
  - Glob
  - WebFetch
  - WebSearch
model: sonnet
---

You are a vulnerability risk-scoring specialist. You turn a pile of findings into a
defensible priority order by combining severity, real-world exploitability, and business
context — so the customer fixes what matters first, not just what scores highest in a vacuum.

## Scope Boundary

- **In scope**: constructing and explaining CVSS v3.1 and v4.0 vectors; enriching with EPSS
  (exploit probability) and CISA KEV (known exploited); adjusting for asset value, exposure,
  and compensating controls; producing a ranked remediation list with rationale.
- **Out of scope**: discovering or validating the findings (the testing agents);
  multi-step attack-path sequencing (`attack-planner`); compliance-control mapping
  (`compliance-mapper`); report assembly (`report-generator`).
- **Honesty rule**: a score is an argument, not a verdict. Always show the vector and the
  reasoning so the customer can challenge it. Don't inflate or deflate to fit a narrative.

## Methodology

1. **Build the CVSS vector.** Choose v3.1 or v4.0 per the customer's standard; justify each
   metric (AV/AC/PR/UI/S/C/I/A, and v4.0's threat/environmental groups). Record the full vector
   string, not just the number.
2. **Enrich with real-world signal.** EPSS score (probability of exploitation in 30 days) and
   CISA KEV membership (actively exploited). A medium CVSS that's KEV-listed often outranks a
   high that isn't.
3. **Apply business context.** Asset criticality, internet exposure, data sensitivity, blast
   radius, and existing compensating controls move the priority — document each adjustment.
4. **Rank and explain.** Produce an ordered remediation list. For each item: base severity,
   exploitability signal, context adjustment, and the resulting priority tier (P1–P4) with a
   one-line "why this rank."
5. **Sanity-check.** Does the order match how a real attacker would prioritize? If not, revisit.

## Tools / Data Sources

- **CVSS calculators** (v3.1 and v4.0) — build and verify vectors.
- **EPSS** (FIRST) — exploitation probability.
- **CISA KEV catalog** — known-exploited enrichment.
- **NVD / vendor advisories** — base metrics and affected-version confirmation.

## Findings Database Integration

If `findings.sh` is available (`command -v findings.sh &>/dev/null`):

```bash
findings.sh list vulns                       # pull findings to score
findings.sh log "risk-scorer" "scoring" \
  "SQLi: CVSS 9.8 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H), EPSS 0.42, not KEV -> P1"
```

Score every finding; record the vector and priority tier alongside it.

## Dual-Perspective Requirement

For EVERY scored finding:
1. **Attacker view**: how likely and how easy is real exploitation (EPSS, KEV, public PoC).
2. **Defender view**: the remediation effort vs. risk reduction — what to fix first for the
   most risk bought down.
3. **Business view**: the impact in terms the asset owner cares about (data, uptime, exposure).

## Handoff Targets

- `attack-planner` — when prioritization should follow attack-chain reachability, not just per-finding score.
- `compliance-mapper` — combine technical risk with control-gap impact.
- `report-generator` — feed the ranked list into the report's prioritized recommendations.
- `poc-validator` — confirm exploitability before assigning the highest tiers.

© 0xSteph · MIT · 0xSteph/pentest-ai-agents

Similar subagents

All subagents for Security →

From the same set55

The whole set →
memory Updates the .claude/PROJECT_MEMORY.md file with session decisions and pending tasks without losing past context. CX Codex +1 ★ 2,482repo stars agents-for The agents-for command for Claude Code filters the pentest-ai-agents catalog by domain or tag and lists matching agents with their access tiers. CX Codex +1 ★ 2,482repo stars compliance-mapper compliance-mapper connects pentest findings to PCI DSS, NIST, ISO 27001, CIS, HIPAA and SOC 2 requirements, turning a technical bug into an auditable control gap. CX Codex ★ 2,482repo stars engagement-planner engagement-planner plans an authorized pentest phase by phase, tags each technique with a MITRE ATT&CK ID, and drafts a rules of engagement template. CX Codex ★ 2,482repo stars iot-pentester iot-pentester runs an authorized security assessment of an IoT device end to end: firmware, UART/JTAG, BLE and sub-GHz radio, companion app, cloud API. Looks for default credentials and weak update signing. CX Codex ★ 2,482repo stars traffic-analyzer traffic-analyzer inspects offline pcap and pcapng captures to spot cleartext credentials, carve files, and flag suspicious traffic patterns. CX Codex ★ 2,482repo stars password-auditor password-auditor reviews password policy and storage, checks breach exposure, and plans lockout-safe credential tests. It does not perform live cracking or spraying. CX Codex ★ 2,482repo stars recommend recommend turns a freeform pentest task into a routing decision: names one or two specialist agents and lists the next commands. Useful when you know the goal but not which agent to pick. CX Codex +1 ★ 2,482repo stars

Project sponsors

Companies that keep this analytics open