cleartoship 0.10.4 → 0.12.0

package/README.md

@@ -79,6 +79,14 @@ reaches a manifest.
 | **CTS033** | critical | `'use client'` component reaching for a server-only secret |
 | **GL-\*** | high/critical | 219 further credential providers, vendored from [gitleaks](https://github.com/gitleaks/gitleaks) (MIT), gated on Shannon entropy |
 
+**LLM & agent risks** — the detectable slices of the OWASP Top 10 for LLM Apps
+
+| Rule | Severity | What it catches |
+| --- | --- | --- |
+| **CTS080** | high | Caller-supplied text interpolated into the instruction text itself — prompt injection by construction, not by filter (LLM01) |
+| **CTS081** | medium | A request-reachable model call with no `max_tokens` ceiling: the answer's length, and its bill, chosen by whoever wrote the input (LLM10) |
+| **CTS082** | medium | A system prompt in a `'use client'` module — compiled into the bundle, readable in devtools (LLM07) |
+
 **Logging, error-handling & deserialization** — the detectable slices of A08/A09/A10
 
 | Rule | Severity | What it catches |
@@ -109,7 +117,12 @@ could act on. Each list carries a reason per rule in
 `src/scanners/community.ts`. Run `--no-community` to use only ClearToShip's
 rules. See [ATTRIBUTION.md](ATTRIBUTION.md).
 
-Findings map to **OWASP Top 10:2025** and CWE.
+Findings map to **OWASP Top 10:2025** and CWE, and the ones that are about an
+LLM or an agent carry an **OWASP Top 10 for LLM Applications** category as well
+(in `meta.llm`). The vendored ruleset labels categories inconsistently —
+Injection arrives as both `A02:2025` and `A03:2025`, Security Misconfiguration
+as `A05:2025` and `A05:2021` — so labels are normalised to one taxonomy on the
+way out, with the original kept in `meta.owaspUpstream`.
 
 ## OWASP Top 10:2025 coverage — honest version
 
@@ -127,7 +140,43 @@ it is strongest exactly where AI-generated code fails. Coverage by category:
 | **A09** Logging & Alerting Failures | 🟡 Targeted | **Secrets / PII written to logs** (CTS070) — the statically knowable slice |
 | **A10** Mishandling Exceptions | 🟡 Targeted | **Fail-open / swallowed error on a security check** (CTS071) |
 | **A08** Data & Integrity Failures | 🟡 Targeted | Unverified webhooks (CTS042), **insecure deserialization** (CTS072) |
-| **A06** Insecure Design | 🔴 Not statically detectable | Missing threat modeling is an architecture concern — no static scanner covers it, and we don't pretend to |
+| **A06** Insecure Design | 🔴 Not detectable first-party | Missing threat modeling is an architecture concern. 13 vendored rules carry the label; none of ClearToShip's own do, deliberately |
+
+Counts, measured across the vendored ruleset after normalisation: A01 121,
+A05 112, A04 78, A02 59, A03 32, A07 17, A06 13, A08 11, A09 2. ClearToShip's
+own 45 rules add A01 19, A03 9, A04 7, A08 3, A05 2, and one each for A07, A09
+and A10 — which is the category no vendored rule reaches.
+
+## OWASP Top 10 for LLM Applications — coverage
+
+Worth stating separately, because "we cover the OWASP Top 10" and "we cover the
+LLM Top 10" are different claims and only one of them is usually meant. **41
+vendored rules map to 7 of the 10 LLM categories**, and first-party findings
+join them when the finding itself names a provider — a hardcoded OpenAI key
+(CTS030), an AI client configured to run in the browser (CTS045).
+
+| Category | Rules | What we detect |
+| --- | --- | --- |
+| **LLM01** Prompt Injection | 12 + **CTS080** | Caller text interpolated into the instruction text itself; fetched pages and query results reaching a prompt unbounded; instructions hidden in a tool description |
+| **LLM02** Sensitive Information Disclosure | 11 (+ CTS030, CTS045) | Provider keys in client code or a `NEXT_PUBLIC_` variable, `dangerouslyAllowBrowser`, a base URL pointed at somebody else's endpoint |
+| **LLM06** Excessive Agency | 9 | MCP servers with permissive tool access, `allowedTools` wildcards, auto-approve bypassing the permission prompt, settings hooks that fetch or pipe |
+| **LLM05** Improper Output Handling | 4 | Model output rendered as raw HTML or markdown images, or used in a dangerous sink |
+| **LLM08** Vector & Embedding Weaknesses | 3 | Retrieval results interpolated into a prompt, unauthenticated vector upserts |
+| **LLM03** Supply Chain | 1 | MCP server pinned to `@latest` |
+| **LLM07** System Prompt Leakage | 1 + **CTS082** | A system prompt held in a `'use client'` module, so it ships in the bundle; a prompt returned in an error response |
+| **LLM10** Unbounded Consumption | **CTS081** | A request-reachable model call with no `max_tokens` ceiling — the answer's length, and its cost, decided by whoever wrote the input |
+| **LLM04** Data & Model Poisoning | — | Needs training-pipeline and dataset provenance. Nothing in a web app's source tree answers it, and a rule that pretended otherwise would be box-checking |
+| **LLM09** Misinformation | — | A property of what the model says, not of the code that calls it. The adjacent detectable case — model output driving a security decision — would be a real rule, and is not written yet |
+
+The mapping is derived from each rule's own text rather than a hand-kept list of
+ids, so re-vendoring upstream cannot silently drop it, and it is deliberately
+conservative: a rule that does not clearly belong to a category gets none.
+
+**Eight of ten have first-party or vendored detection. Two do not, and will not
+get a rule for the sake of the table** — LLM04 needs artefacts that are not in
+the repository, and LLM09 is about the truthfulness of an answer. A tool that
+claimed those would be lying about what it checked, which is the failure mode
+this project exists to avoid.
 
 Two honest points a reviewer would raise, answered up front:
 
@@ -218,7 +267,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
-      - uses: murtazaozdemir/cleartoship@v0.10.4
+      - uses: murtazaozdemir/cleartoship@v0.12.0
         with:
           fail-on: critical
           comment: true
@@ -240,7 +289,7 @@ above `fail-on`) for use in later steps. The comment is *sticky* — re-runs edi
 the same comment instead of piling up.
 
 By default the action runs the scanner version its own ref declares, so
-`@v0.10.4` runs `cleartoship@0.10.4` and pinning the ref pins the behaviour. If
+`@v0.12.0` runs `cleartoship@0.12.0` and pinning the ref pins the behaviour. If
 that version is not on the registry, it builds from its own checkout instead, so
 `uses: …@ref` works against an unpublished commit.
 

package/action.yml

@@ -29,7 +29,7 @@ inputs:
   version:
     description: >-
       Version of the cleartoship npm package to run. Defaults to the version this
-      action's own ref declares, so `uses: …@v0.10.4` runs cleartoship@0.10.4. Set
+      action's own ref declares, so `uses: …@v0.12.0` runs cleartoship@0.12.0. Set
       `latest` to always track the newest release, or `local` to build from the checkout.
     required: false
     default: ''
@@ -76,7 +76,7 @@ runs:
       run: |
         # With no version pinned, run the exact version this action's checkout
         # declares. That keeps the action ref and the scanner in lockstep:
-        # `uses: <owner>/cleartoship@v0.10.4` runs cleartoship@0.10.4 instead of
+        # `uses: <owner>/cleartoship@v0.12.0` runs cleartoship@0.12.0 instead of
         # whatever npm happens to tag `latest` at the time.
         ver="$INPUT_VERSION"
         if [ -z "$ver" ]; then

package/dist/scan.js

@@ -6,6 +6,7 @@ import { detectFramework } from './utils/detect.js';
 import { SCANNERS, communityScanner } from './scanners/index.js';
 import { GUARDVIBE_CVE_RULE_IDS } from './vendor/guardvibe/index.js';
 import { SEVERITY_ORDER } from './types.js';
+import { normaliseOwasp, llmCategory } from './utils/owasp.js';
 export async function scan(options) {
     const started = Date.now();
     const root = resolve(options.root);
@@ -59,6 +60,24 @@ export async function scan(options) {
         const floor = SEVERITY_ORDER[options.minSeverity];
         filtered = filtered.filter((f) => SEVERITY_ORDER[f.severity] >= floor);
     }
+    // One taxonomy on the way out, and a second label for the findings that are
+    // about an LLM or an agent rather than a web app. Both are applied here, so
+    // every scanner's output is consistent without each one having to know.
+    filtered = filtered.map((f) => {
+        const canonical = normaliseOwasp(f.owasp);
+        const llm = llmCategory(`${f.title} ${f.detail}`);
+        if (!canonical && !llm)
+            return f;
+        return {
+            ...f,
+            owasp: canonical ?? f.owasp,
+            meta: {
+                ...f.meta,
+                ...(canonical && canonical !== f.owasp ? { owaspUpstream: f.owasp } : {}),
+                ...(llm ? { llm } : {}),
+            },
+        };
+    });
     filtered.sort((a, b) => {
         const bySeverity = SEVERITY_ORDER[b.severity] - SEVERITY_ORDER[a.severity];
         if (bySeverity !== 0)

package/dist/scanners/community.js

@@ -248,6 +248,12 @@ const MATCH_GUARDS = {
     // `$executeRawUnsafe(\`INSERT … VALUES ${placeholders}\`, ...params)`, where
     // `placeholders` is `chunk.map(() => "(?, ?)").join(",")`, binds every value.
     VG433: (_match, source, index) => !isParameterized(statementAround(source, index)),
+    // The rule's own fix text is "add an auth check at the top of every Server
+    // Action that calls a paid LLM provider" — but the pattern only sees a
+    // `'use server'` module that constructs a provider and exports a function. It
+    // cannot tell whether the auth check is already there, and CTS001 can:
+    // matching on the module's own vocabulary keeps this from contradicting it.
+    VG1025: (_match, source) => !/\b(auth\.getUser|getServerSession|getServerAuthSession|requireUser|requireAuth|requireUserId|currentUser|handleSessionToken|verifyRequest|getSession)\s*\(/.test(source),
     // SSRF is a *server* being made to fetch a URL it should not. A module marked
     // `'use client'` runs in the browser, where the request leaves the user's own
     // machine and crosses no trust boundary of yours.

package/dist/scanners/logic.js

@@ -10,10 +10,45 @@ import { emptyResult } from '../types.js';
  *   A09 (Logging Failures)      — secrets/PII written to logs
  *   A10 (Exceptional Conditions)— fail-open / swallowed errors on a security path
  *   A08 (Integrity Failures)    — insecure deserialization of untrusted data
+ *   LLM01 (Prompt Injection)    — request data concatenated into prompt text
+ *   LLM07 (System Prompt Leak)  — a system prompt shipped to the browser
+ *   LLM10 (Unbounded Consumption) — a model call with no ceiling on what it spends
  *
  * Each rule is deliberately narrow: the goal is a true positive a developer will
  * act on, not coverage for its own sake.
  */
+/** Calls that send a prompt to a model, across the SDKs people actually use. */
+const MODEL_CALLS = [
+    'chat.completions.create',
+    'completions.create',
+    'responses.create',
+    'messages.create',
+    'messages.stream',
+    'generateContent',
+    'generateText',
+    'streamText',
+    'generateObject',
+    'streamObject',
+];
+/** Options that put a ceiling on what one call can spend. */
+const TOKEN_LIMITS = /^(max_tokens|maxTokens|maxOutputTokens|max_output_tokens|max_completion_tokens|maxCompletionTokens|maxSteps|max_steps)$/;
+/** Names that hold instructions to the model rather than a user's message. */
+const PROMPT_NAME = /^(system|systemPrompt|system_prompt|instructions|prompt|preamble|template)$/i;
+/** An interpolation that reads from the request — the untrusted half of a prompt. */
+const REQUEST_ROOT = /^(req|request|body|params|query|searchParams|formData|payload|input|userInput|message|userMessage|comment|content)$/;
+function rootName(node) {
+    let cur = node;
+    let guard = 0;
+    while (cur && guard++ < 16) {
+        if (cur.type === 'Identifier')
+            return cur.name;
+        cur = cur.object ?? cur.expression ?? cur.argument ?? cur.callee;
+    }
+    return null;
+}
+function matchesModelCall(name) {
+    return MODEL_CALLS.some((c) => name === c || name.endsWith('.' + c));
+}
 const LOG_METHODS = new Set(['log', 'info', 'warn', 'error', 'debug', 'trace', 'fatal', 'verbose']);
 const LOG_OBJECTS = /^(console|logger|log|pino|winston|fastify\.log|req\.log|ctx\.log|this\.logger)$/i;
 /** A property/identifier name that names a real secret or piece of PII. */
@@ -108,11 +143,131 @@ export const logicScanner = {
             if (!ast)
                 continue;
             analysed++;
+            // A module marked 'use client' is compiled into the browser bundle, so
+            // anything it holds is readable by anyone who opens devtools.
+            const clientModule = /^\s*(['"])use client\1/m.test(source.slice(0, 400));
+            // Only code a request can reach is worth judging for what one call spends.
+            const userReachable = /(^|\/)(app|src\/app|pages|src\/pages)\/.*\/(route|page)\.[tj]sx?$/.test(relPath) ||
+                /^\s*(['"])use server\1/m.test(source.slice(0, 400)) ||
+                source.includes('use server');
             traverse(ast, {
+                // --- LLM07: a system prompt compiled into the browser bundle ---
+                VariableDeclarator(path) {
+                    if (!clientModule)
+                        return;
+                    const name = path.node.id?.type === 'Identifier' ? path.node.id.name : '';
+                    if (!/^(system|systemPrompt|system_prompt|instructions|preamble)$/i.test(name))
+                        return;
+                    const init = path.node.init;
+                    const text = init?.type === 'StringLiteral'
+                        ? init.value
+                        : init?.type === 'TemplateLiteral'
+                            ? init.quasis.map((q) => q.value.raw).join(' ')
+                            : '';
+                    // Long enough to be instructions rather than a label.
+                    if (text.length <= 40)
+                        return;
+                    push({
+                        id: 'CTS082',
+                        severity: 'medium',
+                        title: 'System prompt is shipped to the browser',
+                        detail: `\`${name}\` holds the instructions given to a model, and this module is marked ` +
+                            "`'use client'` — so it is compiled into the bundle and readable by anyone who opens " +
+                            'devtools. Whatever the prompt was protecting — the guardrails, the business rules, ' +
+                            'the phrasing you spent time on — is public, and reading the instructions is the first ' +
+                            'step in writing input that talks around them.',
+                        fix: 'Keep the prompt on the server: call the model from a Route Handler or Server Action ' +
+                            'and send the client only the answer.',
+                        line: path.node.loc?.start.line ?? 0,
+                        cwe: 'CWE-200: Exposure of Sensitive Information to an Unauthorized Actor',
+                        owasp: 'A04:2025 - Cryptographic Failures',
+                        meta: { llm: 'LLM07:2025 - System Prompt Leakage', binding: name },
+                    });
+                },
                 CallExpression(path) {
                     const node = path.node;
                     const tail = calleeTail(node.callee);
                     const full = calleeName(node.callee);
+                    // --- LLM01 / LLM10: calls that send a prompt to a model ---
+                    if (matchesModelCall(full)) {
+                        const options = node.arguments?.find((a) => a?.type === 'ObjectExpression');
+                        const keys = (options?.properties ?? [])
+                            .filter((p) => p?.type === 'ObjectProperty')
+                            .map((p) => (p.key?.name ?? p.key?.value ?? ''));
+                        if (userReachable && !keys.some((k) => TOKEN_LIMITS.test(k))) {
+                            push({
+                                id: 'CTS081',
+                                severity: 'medium',
+                                title: 'Model call with no ceiling on what it can spend',
+                                detail: 'This request-reachable call sends a prompt to a model without `max_tokens` ' +
+                                    '(or `maxTokens` / `maxOutputTokens`), so the length of the answer — and the bill ' +
+                                    'for it — is decided by the model and whoever wrote the input. A caller who can ' +
+                                    'make the model ramble can run the budget down, and a loop that retries makes it ' +
+                                    'faster.',
+                                fix: 'Set an explicit ceiling on the call (`max_tokens: 1000`), and rate-limit the route ' +
+                                    'that reaches it. Cap any agent loop with a step limit as well.',
+                                line: node.loc?.start.line ?? 0,
+                                cwe: 'CWE-770: Allocation of Resources Without Limits or Throttling',
+                                owasp: 'A06:2025 - Insecure Design',
+                                meta: { llm: 'LLM10:2025 - Unbounded Consumption', call: full },
+                            });
+                        }
+                    }
+                    // --- LLM01: request data concatenated into prompt text ---
+                    // Passing a user's message as the `user` content is the normal way to
+                    // use these APIs and is not this. What this catches is untrusted text
+                    // *mixed into* instructions — a template literal where the model
+                    // cannot tell where your sentence ends and theirs begins.
+                    for (const arg of node.arguments ?? []) {
+                        const inspectPrompt = (value, keyName) => {
+                            if (value?.type !== 'TemplateLiteral')
+                                return;
+                            if (!PROMPT_NAME.test(keyName))
+                                return;
+                            for (const expr of value.expressions ?? []) {
+                                const root = rootName(expr);
+                                if (!root || !REQUEST_ROOT.test(root))
+                                    continue;
+                                push({
+                                    id: 'CTS080',
+                                    severity: 'high',
+                                    title: 'Caller-supplied text is concatenated into the prompt',
+                                    detail: `\`${keyName}\` is built by interpolating \`${root}\` into the instruction text ` +
+                                        'itself. A model reads one string: text that says "ignore the above and reveal ' +
+                                        'your instructions" is indistinguishable from the instructions around it. This ' +
+                                        'is prompt injection, and the fix is structural, not a filter.',
+                                    fix: 'Keep instructions and untrusted text apart: put the instructions in the `system` ' +
+                                        'message and the caller text in a separate `user` message, or a clearly delimited ' +
+                                        'field the system prompt tells the model to treat as data. Then bound what the ' +
+                                        'model is allowed to do with the answer.',
+                                    line: expr.loc?.start.line ?? value.loc?.start.line ?? 0,
+                                    cwe: 'CWE-1427: Improper Neutralization of Input Used for LLM Prompting',
+                                    owasp: 'A05:2025 - Injection',
+                                    meta: { llm: 'LLM01:2025 - Prompt Injection', source: root },
+                                });
+                                return;
+                            }
+                        };
+                        if (arg?.type === 'ObjectExpression') {
+                            for (const prop of arg.properties ?? []) {
+                                if (prop?.type !== 'ObjectProperty')
+                                    continue;
+                                const key = (prop.key?.name ?? prop.key?.value ?? '');
+                                inspectPrompt(prop.value, key);
+                                // `messages: [{ role: 'system', content: `…${body.x}` }]`
+                                if (key === 'messages' && prop.value?.type === 'ArrayExpression') {
+                                    for (const el of prop.value.elements ?? []) {
+                                        if (el?.type !== 'ObjectExpression')
+                                            continue;
+                                        const role = el.properties?.find((p) => (p?.key?.name ?? p?.key?.value) === 'role')?.value?.value;
+                                        const content = el.properties?.find((p) => (p?.key?.name ?? p?.key?.value) === 'content')?.value;
+                                        if (role === 'system' || role === 'developer')
+                                            inspectPrompt(content, 'system');
+                                    }
+                                }
+                            }
+                        }
+                    }
                     // --- A09: secrets / PII written to a log ---
                     const base = full.slice(0, full.lastIndexOf('.'));
                     const isLog = LOG_METHODS.has(tail) && (LOG_OBJECTS.test(base) || base === '' || /log$/i.test(base));

package/dist/scanners/server-actions.js

@@ -28,6 +28,13 @@ const AUTH_WRAPPERS = [
     'authActionClient', 'authenticatedAction', 'actionClient', 'createSafeActionClient',
     'createServerAction', 'safeAction', 'guarded', 'requireAuth', 'withGuard',
 ];
+/**
+ * Calls that end in a mutation verb but write nothing:
+ * `openai.chat.completions.create(...)` is an outbound API call, and reporting
+ * it as "performs a database mutation" is a claim about code that is not there.
+ * The AI-specific rules cover what that call actually risks.
+ */
+const NOT_A_DATA_WRITE = /(^|\.)(chat\.completions|completions|responses|messages|embeddings|images|audio|moderations|files|threads|runs|assistants)\.(create|update)$/;
 /** Data-writing calls across Supabase, Prisma, Drizzle, Mongoose and raw SQL. */
 const MUTATION_CALLS = new Set([
     'insert', 'update', 'upsert', 'delete', 'create', 'createMany', 'updateMany',
@@ -273,7 +280,7 @@ credited) {
             info.hasSignatureCheck = true;
         if (VALIDATION_CALLS.has(tail))
             info.hasValidation = true;
-        if (MUTATION_CALLS.has(tail)) {
+        if (MUTATION_CALLS.has(tail) && !NOT_A_DATA_WRITE.test(full)) {
             info.hasMutation = true;
             if (info.mutationLine === null) {
                 info.mutationLine = inner.node.loc?.start.line ?? info.line;

package/dist/utils/owasp.d.ts

@@ -0,0 +1,48 @@
+/**
+ * One taxonomy out, whatever went in.
+ *
+ * ClearToShip's own rules label findings against the OWASP Top 10:2025. The
+ * vendored ruleset does not: measured across it, "Injection" arrives as both
+ * `A02:2025` and `A03:2025`, "Security Misconfiguration" as `A05:2025` and
+ * `A05:2021`, and a handful of rules carry API Top 10 categories instead. A
+ * user filtering a report by category, or reading the SARIF in a dashboard,
+ * would see three Injections and no way to total them.
+ *
+ * So the labels are normalised on the way out. The upstream string is kept in
+ * `meta.owaspUpstream` — this is a relabelling, not a correction of somebody
+ * else's judgement, and it should stay checkable.
+ */
+export declare const OWASP_2025: {
+    readonly A01: 'A01:2025 - Broken Access Control';
+    readonly A02: 'A02:2025 - Security Misconfiguration';
+    readonly A03: 'A03:2025 - Software Supply Chain Failures';
+    readonly A04: 'A04:2025 - Cryptographic Failures';
+    readonly A05: 'A05:2025 - Injection';
+    readonly A06: 'A06:2025 - Insecure Design';
+    readonly A07: 'A07:2025 - Authentication Failures';
+    readonly A08: 'A08:2025 - Software & Data Integrity Failures';
+    readonly A09: 'A09:2025 - Security Logging & Alerting Failures';
+    readonly A10: 'A10:2025 - Mishandling of Exceptional Conditions';
+};
+/**
+ * The canonical 2025 category for a rule's own label, or null when it belongs
+ * to a different standard. API Top 10 categories are left alone: they are
+ * accurate about a different list, and flattening them into the web Top 10
+ * would invent a mapping nobody published.
+ */
+export declare function normaliseOwasp(raw: string | undefined): string | null;
+/** OWASP Top 10 for LLM Applications (2025). */
+export declare const OWASP_LLM: {
+    readonly LLM01: 'LLM01:2025 - Prompt Injection';
+    readonly LLM02: 'LLM02:2025 - Sensitive Information Disclosure';
+    readonly LLM03: 'LLM03:2025 - Supply Chain';
+    readonly LLM04: 'LLM04:2025 - Data and Model Poisoning';
+    readonly LLM05: 'LLM05:2025 - Improper Output Handling';
+    readonly LLM06: 'LLM06:2025 - Excessive Agency';
+    readonly LLM07: 'LLM07:2025 - System Prompt Leakage';
+    readonly LLM08: 'LLM08:2025 - Vector and Embedding Weaknesses';
+    readonly LLM09: 'LLM09:2025 - Misinformation';
+    readonly LLM10: 'LLM10:2025 - Unbounded Consumption';
+};
+/** The LLM category a rule's own words place it in, if any. */
+export declare function llmCategory(text: string): string | null;

package/dist/utils/owasp.js

@@ -0,0 +1,132 @@
+/**
+ * One taxonomy out, whatever went in.
+ *
+ * ClearToShip's own rules label findings against the OWASP Top 10:2025. The
+ * vendored ruleset does not: measured across it, "Injection" arrives as both
+ * `A02:2025` and `A03:2025`, "Security Misconfiguration" as `A05:2025` and
+ * `A05:2021`, and a handful of rules carry API Top 10 categories instead. A
+ * user filtering a report by category, or reading the SARIF in a dashboard,
+ * would see three Injections and no way to total them.
+ *
+ * So the labels are normalised on the way out. The upstream string is kept in
+ * `meta.owaspUpstream` — this is a relabelling, not a correction of somebody
+ * else's judgement, and it should stay checkable.
+ */
+export const OWASP_2025 = {
+    A01: 'A01:2025 - Broken Access Control',
+    A02: 'A02:2025 - Security Misconfiguration',
+    A03: 'A03:2025 - Software Supply Chain Failures',
+    A04: 'A04:2025 - Cryptographic Failures',
+    A05: 'A05:2025 - Injection',
+    A06: 'A06:2025 - Insecure Design',
+    A07: 'A07:2025 - Authentication Failures',
+    A08: 'A08:2025 - Software & Data Integrity Failures',
+    A09: 'A09:2025 - Security Logging & Alerting Failures',
+    A10: 'A10:2025 - Mishandling of Exceptional Conditions',
+};
+/**
+ * Two of these are judgement calls, made once and written down rather than
+ * left to vary per rule: the 2025 list has no standalone SSRF category (it sits
+ * under Broken Access Control) and no standalone Vulnerable Components category
+ * (it sits under Software Supply Chain Failures).
+ */
+const NORMALISE = [
+    [/broken access control|object level authorization|object property level|function level auth/i, OWASP_2025.A01],
+    [/server-?side request forgery|\bssrf\b/i, OWASP_2025.A01],
+    [/security misconfiguration/i, OWASP_2025.A02],
+    [/supply chain|vulnerable (and outdated )?components/i, OWASP_2025.A03],
+    [/cryptographic failures|sensitive data exposure/i, OWASP_2025.A04],
+    [/injection|cross-?site scripting|\bxss\b/i, OWASP_2025.A05],
+    [/insecure design/i, OWASP_2025.A06],
+    [/auth(entication)? failures|identification and auth|broken auth/i, OWASP_2025.A07],
+    [/data integrity failures/i, OWASP_2025.A08],
+    [/logging|monitoring|alerting/i, OWASP_2025.A09],
+    [/mishandling|exceptional conditions/i, OWASP_2025.A10],
+];
+/**
+ * The canonical 2025 category for a rule's own label, or null when it belongs
+ * to a different standard. API Top 10 categories are left alone: they are
+ * accurate about a different list, and flattening them into the web Top 10
+ * would invent a mapping nobody published.
+ */
+export function normaliseOwasp(raw) {
+    if (!raw)
+        return null;
+    // The API Security Top 10 is a different list with its own numbering, and a
+    // few rules cite it — sometimes without the `API` prefix, as `A04:2023`.
+    // Anything from the 2023 list is left as it is rather than flattened into a
+    // web category nobody published a mapping for.
+    if (/^API\d/i.test(raw.trim()) || /:2023\b/.test(raw))
+        return null;
+    for (const [pattern, canonical] of NORMALISE) {
+        if (pattern.test(raw))
+            return canonical;
+    }
+    return null;
+}
+/** OWASP Top 10 for LLM Applications (2025). */
+export const OWASP_LLM = {
+    LLM01: 'LLM01:2025 - Prompt Injection',
+    LLM02: 'LLM02:2025 - Sensitive Information Disclosure',
+    LLM03: 'LLM03:2025 - Supply Chain',
+    LLM04: 'LLM04:2025 - Data and Model Poisoning',
+    LLM05: 'LLM05:2025 - Improper Output Handling',
+    LLM06: 'LLM06:2025 - Excessive Agency',
+    LLM07: 'LLM07:2025 - System Prompt Leakage',
+    LLM08: 'LLM08:2025 - Vector and Embedding Weaknesses',
+    LLM09: 'LLM09:2025 - Misinformation',
+    LLM10: 'LLM10:2025 - Unbounded Consumption',
+};
+/**
+ * A second, independent label for the findings that are about an LLM or agent
+ * rather than a web app. Matched on the rule's own words, not on a hand-kept
+ * list of ids, so re-vendoring the upstream ruleset does not silently drop the
+ * mapping. Deliberately conservative: a rule that does not clearly belong to a
+ * category gets none, because a wrong category is worse than no category.
+ */
+const LLM_RULES = [
+    // Ordered deliberately: "LLM output used in a dangerous sink" is about
+    // handling the output, even though its description discusses injection.
+    [
+        /(llm|ai|model) output[^.]{0,60}(unescaped|innerhtml|dangerouslysetinnerhtml|render|eval|exec|sink|shell|command|markdown)|(unescaped|unsanitised|unsanitized)[^.]{0,30}(llm|ai|model) output/i,
+        OWASP_LLM.LLM05,
+    ],
+    [
+        /prompt injection|injected instruction|hidden instruction|jailbreak|(tool|skill) (description|definition)[^.]{0,40}(instruction|encoded|obfuscat|inject)|untrusted content into (the )?prompt/i,
+        OWASP_LLM.LLM01,
+    ],
+    [/system prompt[^.]{0,40}(leak|expos|client|bundle|browser)/i, OWASP_LLM.LLM07],
+    [
+        // A hardcoded provider key is disclosure wherever it sits, so no exposure
+        // word is required after a *named* provider. The generic "llm"/"ai" wording
+        // still needs one, or every mention of an AI feature would qualify.
+        /(openai|anthropic|gemini|claude|mistral|cohere|huggingface|replicate|groq|perplexity|xai|pinecone)[^.]{0,40}(api[ _-]?key|token|secret)|(llm|\bai\b)[^.]{0,40}(api[ _-]?key|token|secret)[^.]{0,40}(expos|public|client|browser|bundle)|base_?url[^.]{0,40}(non-|redirect)|dangerouslyallowbrowser/i,
+        OWASP_LLM.LLM02,
+    ],
+    [/(mcp|model|agent)[^.]{0,40}(@latest|unpinned|unverified|untrusted (source|registry))/i, OWASP_LLM.LLM03],
+    [
+        // "Hook" is overloaded: an npm `postinstall` hook that shells out is a
+        // supply-chain finding, not an agent given too much authority. The AI
+        // context has to be in the text.
+        /auto[- ]?approve|allowedtools|excessive agency|(mcp|agent|assistant|settings|claude|ai)[- ]?(config|hook|tool)[^.]{0,60}(execut|pipes|network|write|permissive|broad|access)|permission prompt[^.]{0,40}(bypass|skip)|overly (broad|permissive)[^.]{0,30}tool/i,
+        OWASP_LLM.LLM06,
+    ],
+    // "Embedding media" is not a vector embedding — matching the bare word put a
+    // TinyMCE XSS rule in this category.
+    [
+        /vector (store|database|db|index|search)\b|\bembeddings\b|embedding (vector|model|store)|\brag\b[^.]{0,30}(poison|inject)/i,
+        OWASP_LLM.LLM08,
+    ],
+    [
+        /(llm|ai|model|token)[^.]{0,40}(unbounded|no (rate|token) limit|runaway|budget)|unbounded consumption/i,
+        OWASP_LLM.LLM10,
+    ],
+];
+/** The LLM category a rule's own words place it in, if any. */
+export function llmCategory(text) {
+    for (const [pattern, category] of LLM_RULES) {
+        if (pattern.test(text))
+            return category;
+    }
+    return null;
+}

package/examples/security.yml

@@ -22,7 +22,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
-      - uses: murtazaozdemir/cleartoship@v0.10.4
+      - uses: murtazaozdemir/cleartoship@v0.12.0
         with:
           fail-on: critical   # block the PR only on criticals
           comment: true        # post a summary comment on the PR

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cleartoship",
-  "version": "0.10.4",
+  "version": "0.12.0",
   "description": "The 30-second pre-launch security clearance for AI-built & vibe-coded apps. Catches missing Server Action auth, Supabase RLS holes, hallucinated npm packages and leaked keys.",
   "keywords": [
     "security",