feat: add git node security --validate-reports

components/git/security.js

@@ -3,6 +3,7 @@ import PrepareSecurityRelease from '../../lib/prepare_security.js';
 import UpdateSecurityRelease from '../../lib/update_security_release.js';
 import SecurityBlog from '../../lib/security_blog.js';
 import SecurityAnnouncement from '../../lib/security-announcement.js';
+import ValidateReports from '../../lib/validate_reports.js';
 
 export const command = 'security [options]';
 export const describe = 'Manage an in-progress security release or start a new one.';
@@ -40,6 +41,52 @@ const securityOptions = {
     describe: 'Request CVEs for a security release',
     type: 'boolean'
   },
+  'validate-reports': {
+    describe: 'Validate triaged HackerOne reports against the Node.js threat model',
+    type: 'boolean'
+  },
+  'validate-reports-format': {
+    choices: ['markdown', 'json'],
+    default: 'markdown',
+    describe: 'Output format for --validate-reports',
+    type: 'string'
+  },
+  'validate-reports-output': {
+    describe: 'Write --validate-reports output to a file instead of stdout',
+    type: 'string'
+  },
+  'validate-reports-limit': {
+    describe: 'Maximum number of triaged reports to validate',
+    type: 'number'
+  },
+  'validate-reports-confirm': {
+    default: true,
+    describe: 'Ask before continuing to the next report after each LLM assessment',
+    type: 'boolean'
+  },
+  'validate-reports-cache': {
+    default: true,
+    describe: 'Reuse cached LLM assessments for the same report, model, and prompt',
+    type: 'boolean'
+  },
+  llm: {
+    choices: ['codex', 'claude', 'copilot'],
+    describe: 'Ask an LLM CLI to assess each triaged report',
+    type: 'string'
+  },
+  'llm-model': {
+    describe: 'Override the LLM model used for command construction and cache identity',
+    type: 'string'
+  },
+  'llm-command': {
+    describe: 'Override the command used for --llm. The report prompt is sent on stdin.',
+    type: 'string'
+  },
+  'node-repo': {
+    default: process.cwd(),
+    describe: 'Node.js checkout path the LLM should use to read SECURITY.md and doc/',
+    type: 'string'
+  },
   'post-release': {
     describe: 'Create the post-release announcement to the given nodejs.org folder',
     type: 'string'
@@ -82,6 +129,9 @@ export function builder(yargs) {
       'git node security --request-cve',
       'Request CVEs for a security release of Node.js based on' +
       ' the next-security-release/vulnerabilities.json'
+    ).example(
+      'git node security --validate-reports',
+      'Validate triaged HackerOne reports against the Node.js threat model'
     ).example(
       'git node security --post-release="../nodejs.org/"',
       'Create the post-release announcement on the Nodejs.org repo'
@@ -119,6 +169,9 @@ export function handler(argv) {
   if (argv['request-cve']) {
     return requestCVEs(cli, argv);
   }
+  if (argv['validate-reports']) {
+    return validateReports(cli, argv);
+  }
   if (argv['post-release']) {
     return createPostRelease(cli, argv);
   }
@@ -157,6 +210,11 @@ async function requestCVEs(cli) {
   return hackerOneCve.requestCVEs();
 }
 
+async function validateReports(cli, argv) {
+  const validator = new ValidateReports(cli, argv);
+  return validator.validate();
+}
+
 async function createPostRelease(cli, argv) {
   const nodejsOrgFolder = argv['post-release'];
   const blog = new SecurityBlog(cli);

docs/git-node.md

@@ -506,6 +506,136 @@ Example:
   git node security --remove-report=12345
 ```
 
+### `git node security --validate-reports`
+
+This command retrieves all triaged HackerOne reports for the Node.js program and
+produces a local validation report for each one. It is intended to help the
+security team review whether a report still looks valid under the Node.js threat
+model and whether the current HackerOne severity/CVSS is consistent with the
+available evidence.
+
+The command uses the existing HackerOne credentials configured in `.ncurc`. It
+does not modify HackerOne reports, labels, comments, or severity. The output is
+only a local triage aid and still requires human review.
+
+```sh
+  git node security --validate-reports
+  git node security --validate-reports --validate-reports-format=json
+  git node security --validate-reports --validate-reports-output=reports.md
+  git node security --validate-reports --llm=codex --node-repo=/path/to/node
+  git node security --validate-reports --llm=codex --llm-model=gpt-5.5
+  git node security --validate-reports --llm=codex --no-validate-reports-confirm
+  git node security --validate-reports --llm=codex --no-validate-reports-cache
+  git node security --validate-reports --llm=claude --node-repo=/path/to/node
+  git node security --validate-reports --llm=copilot --node-repo=/path/to/node
+  git node security --validate-reports --llm=copilot --llm-command="copilot -p"
+```
+
+By default, the command runs a heuristic pass only. The heuristic checks the
+report title, vulnerability information, impact, description, comments, current
+severity, CVSS vector, and weakness metadata for common Node.js security topics.
+It can identify obvious mismatches, such as a CVSS vector whose calculated
+rating does not match the HackerOne rating. Keyword matches are treated only as
+topic hints, not as proof that a report is valid or invalid. This is deliberate:
+HackerOne report text is reporter-controlled, so words like `request smuggling`
+or `permission model` are not enough to make a threat-model decision. The
+heuristic output is deliberately conservative and always leaves threat-model
+validity as `needs-manual-review`.
+
+Use `--llm=<provider>` to ask an LLM CLI to produce a structured assessment for
+each report. Supported providers are `codex`, `claude`, and `copilot`.
+
+When LLM mode is enabled, the command asks before assessing each report and shows
+the report title, current severity, CVSS vector, and weakness. After each LLM
+assessment, it prints a readable summary with:
+
+- the report URL and title
+- the provider and model/cache identity
+- validity under the Node.js threat model
+- whether the current severity is correct
+- current severity/CVSS and suggested severity/CVSS
+- a colored CVSS metric diff when the suggested vector differs
+- CWE
+- confidence from 0 to 100
+- threat model/documentation references used by the model
+- reasoning
+
+Use `--no-validate-reports-confirm` for batch mode without the per-report
+prompts. Use `--validate-reports-limit=<n>` to test the flow against a smaller
+number of reports.
+
+#### LLM prompt and reasoning
+
+The LLM prompt is designed to keep the model anchored to the Node.js threat
+model and local documentation instead of only reasoning from the report text.
+For each report, the prompt instructs the model to:
+
+- read `SECURITY.md` from the Node.js checkout supplied through `--node-repo`
+- inspect relevant files under `doc/` for the affected API or subsystem
+- apply the documented Node.js threat model, including treatment of application
+  code, caller-supplied API inputs, third-party modules, unsupported platforms,
+  and inspector/debugger access
+- decide whether the report is valid under that threat model
+- decide whether the current HackerOne severity/CVSS is correct
+- use reports with the same CWE/weakness as precedent context when available
+- avoid copying precedent blindly when `SECURITY.md` or `doc/` point to a
+  different result
+- return only JSON matching the schema expected by the command
+
+The report payload sent to the model contains the HackerOne report id, title,
+URL, state, current severity, current CVSS vector, weakness metadata, reporter,
+report body fields, comments, heuristic findings, and comparable reports with
+the same weakness. Comparable reports include their title, URL, current
+severity/CVSS, state, and any team summary available through HackerOne. This
+helps the model account for previous team decisions while still checking the
+current report against the threat model.
+
+The model must return these fields:
+
+- `validity`: `valid`, `invalid`, or `needs-more-info`
+- `severity_correct`: boolean
+- `suggested_severity`: `none`, `low`, `medium`, `high`, `critical`, or
+  `informational`
+- `suggested_cvss`: a CVSS vector or `N/A`
+- `cwe`: the best matching CWE
+- `confidence`: a number from 0 to 100
+- `reasoning`: a concise explanation
+- `threat_model_references`: references to `SECURITY.md` and relevant `doc/`
+  material used for the decision
+
+#### LLM commands and cache
+
+The `--llm-command` option can override the default provider command. The prompt
+is sent on stdin and the command must print a JSON object matching the expected
+schema.
+
+The model label is inferred from the local LLM CLI configuration when possible.
+For example, Codex reads `~/.codex/config.toml` and includes
+`model_reasoning_effort` in the cache label. Use `--llm-model` to override the
+provider command model and the cache identity. If the model cannot be inferred,
+the cache entry includes a comment explaining that `default` was used.
+
+Successful LLM assessments are cached locally in
+`.ncu-cache/security-report-validation` using the report, provider, model, and
+prompt as the cache key. Use `--no-validate-reports-cache` to force a fresh LLM
+assessment.
+
+#### Options
+
+| Option | Description |
+| --- | --- |
+| `--validate-reports-format=markdown\|json` | Select the final output format. Defaults to `markdown`. |
+| `--validate-reports-output=<file>` | Write the final output to a file instead of stdout. |
+| `--validate-reports-limit=<n>` | Validate at most `n` triaged reports. Useful for testing the flow. |
+| `--validate-reports-confirm` | Ask before each LLM assessment and before continuing to the next report. Enabled by default. |
+| `--no-validate-reports-confirm` | Disable interactive prompts for batch runs. |
+| `--validate-reports-cache` | Reuse cached successful LLM assessments. Enabled by default. |
+| `--no-validate-reports-cache` | Ignore existing LLM cache entries and do not reuse them. |
+| `--llm=codex\|claude\|copilot` | Ask an LLM CLI to assess each report. |
+| `--llm-model=<model>` | Override the provider model and cache identity. |
+| `--llm-command=<command>` | Override the command used for LLM assessment. The prompt is sent on stdin. |
+| `--node-repo=<path>` | Path to a Node.js checkout containing `SECURITY.md` and `doc/`. Defaults to the current directory. |
+
 ## `git node status`
 
 Return status and information about the current git-node land session. Shows the following information: