What Sekrd’s November 2026 scanner update checks — and what it refuses to flag
Short answer
The November 2026 scanner update adds firm, low-false-positive checks for secrets and misconfigurations that show up in AI-built apps: new provider key formats in the JavaScript bundle, a Supabase sb_secret_ key (it bypasses Row Level Security), a Gemini key used against the paid generativelanguage API, server secret names behind NEXT_PUBLIC_, VITE_, EXPO_PUBLIC_ and similar prefixes, a system prompt in the bundle that contains credentials or an authorization rule (OWASP LLM07:2025), and an exact pin of a React Server Components package affected by CVE-2025-55182. A deep scan adds read-only probes for the Next.js middleware bypass CVE-2025-29927, an MCP endpoint that lists tools with no authentication, GraphQL introspection that names sensitive fields, a public S3 or GCS bucket that lists credential or database objects, and a real .env file or /.git/HEAD. A bare 200, a public image bucket, a Maps API key, or a normal system prompt is not a finding. The update does not send exploit payloads.
Last updated · Sekrd Research
- Free scans analyze the shipped JavaScript. Deep scans add the read-only probes.
- Every new check has a lookalike that must stay silent:
sb_publishable_, a bareAIzaMaps key, GraphQL introspection of ordinary content fields, MCPinitializealone, a public image bucket. - No prompt-injection payloads, no React2Shell request, and no Next.js version guessed from the homepage.
- A leaked
sb_secret_key is a hard-block finding: it maps toservice_role, which bypasses RLS, and a firm hard-block forces the trust score to 0.
Why these checks, and why so few
AI coding tools now call more providers than OpenAI and Anthropic, and they inline whatever environment variable you prefix for the browser. A scanner that flags every AIza string, every GraphQL endpoint, or every system prompt trains founders to ignore it. This update only fires when the evidence is the leak itself: a vendor prefix, a secret name on a client-side env var, a protected page that actually opens, or a file body that is a real env file or a git ref.
Confidence is part of the result. Pattern matches on these new rules are firm. The deep-scan probes are confirmed: they require a positive signal, not a status code. Details and the fix for each finding are on the report. The rules below are the ones that changed.
What a free scan now reads in the bundle
These run on every scan. They only look at JavaScript the site already sent. They do not make extra attack requests.
New AI provider key formats
Critical, firm, CWE-798. The bundle is searched for prefixes that vendors document as secret keys:
- OpenAI service-account keys,
sk-svcacct- - xAI keys,
xai- - Perplexity keys,
pplx- - Fireworks keys,
fw_ - ElevenLabs keys only when the name is there too (
xi-api-keyor an ElevenLabs env assignment). A random 32-character hex string is not an ElevenLabs finding.
Older OpenAI and Anthropic prefixes were already covered. Publishable tokens that are supposed to live in the browser stay quiet.
Supabase sb_secret_
Critical, firm, and a hard block. Supabase’s secret key replaces the legacy service_role JWT and bypasses every Row Level Security policy. The publishable key, sb_publishable_, is designed for browsers and is not reported. See the migration guide.
Gemini key on the paid API
Critical, firm, only when an AIza key sits next to generativelanguage.googleapis.com. A bare AIza key stays a low, tentative note: the same shape is a normal Maps, Firebase, or reCAPTCHA browser key. The cost-exposure check follows the same rule. What to do if you hit this is in the leaked-key guide.
Server secrets behind a client env prefix
High, firm, CWE-798. Build tools inline NEXT_PUBLIC_, VITE_, EXPO_PUBLIC_, NUXT_PUBLIC_, and SvelteKit PUBLIC_ into the bundle. The finding fires when that public name is a server secret (an LLM key, STRIPE_SECRET, SERVICE_ROLE, DATABASE_URL, a webhook secret, and the same class of name). Public-by-design names — anon, publishable, site URL — are not in the list. Comments and placeholders (example, your_, changeme, todo) are skipped. The full rule is in the env-var guide.
System prompt in the bundle
OWASP LLM07:2025 says a system prompt is not a secret and must not be the access-control layer. We follow that. A systemPrompt assigned in client JavaScript is a finding only when the string contains credential material (high, CWE-200) or an authorization rule such as “you are an admin” or “bypass RLS” (medium). A normal product prompt is not reported, and we do not send prompt-injection payloads to your model.
Vulnerable React Server Components pin
Critical, firm, CWE-1395. The shipped artifact must pin react-server-dom-webpack, react-server-dom-parcel, or react-server-dom-turbopack at 19.0.0, 19.1.0, 19.1.1, or 19.2.0 — the versions in CVE-2025-55182 (CVSS 10). Patched revisions 19.0.1, 19.1.2, and 19.2.1 do not match. A sentence on a marketing page that mentions the version does not match either: the check requires a package pin, not prose. It does not send an exploit payload. How to check the lockfile is in the CVE-2025-55182 guide. A later advisory, GHSA-xqc2-xf4c-x3h3 on October 14, 2026, names a wider set of versions. That set is not what this check matches. The write-up is the October 14 release.
What a deep scan probes, still without an exploit
Active testing runs only on a deep scan (Pre-Launch Audit or Continuous Pro). Each probe is read-only, throttled with the rest of active DAST, and silent unless the dangerous condition is actually there.
- Next.js middleware bypass (CVE-2025-29927). Critical, confirmed, CWE-288. Only if the site looks like Next.js, a path such as
/dashboardor/adminis behind a real auth gate, and that same path then returns the app instead of the login page. A homepage, a locale redirect, a WAF block, or a server that ignores the internal header produces nothing. We do not guess a Next.js version from the homepage. Background and the version check are in the CVE-2025-29927 guide. - Unauthenticated MCP
tools/list. High, or critical if a tool name looks like SQL, shell, file, or admin. CWE-306, the same class as OWASP LLM06:2025 Excessive Agency.initializealone is not a finding: the MCP auth spec allows an unauthenticated initialize followed by a 401 ontools/list. The check does not call any tool. See the MCP checklist. - GraphQL introspection of sensitive fields. Medium, confirmed, CWE-200. Anonymous introspection is reported only when the schema names fields such as
password,api_key,service_role, or a card number. A public content API that introspectsPost.titleis not a finding. - Public S3 or GCS buckets. Critical or high, confirmed, CWE-552. We only request a listing for a bucket URL already present in the page, and only report it when the listing names a credential or database object (
.env,.pem,.sql, a dump, a backup archive). A bucket of images is not a finding. - Exposed env files and
/.git/HEAD. Critical, confirmed./.env,/.env.local,/.env.production, and/.env.developmentmust come back as a realKEY=valuefile with a secret-shaped name, not an HTML page that mentionsAPI_KEY./.git/HEADmust be a git ref (ref: refs/…) or a commit SHA, which an SPA fallback cannot fake.
What we deliberately left out
- Prompt-injection strings, and “the model’s output is written to
innerHTML”. Both fire on ordinary apps. - A React2Shell exploit request, and a Next.js version inferred from the HTML.
- GraphQL introspection by itself, MCP
initializeby itself, and public image buckets. - More missing-header findings, OAuth PKCE heuristics, and retire.js-style library fingerprinting. A version string has to be unambiguous before we will name a CVE from it.
Each new check has a regression case (the lookalike that must stay silent) and an anti-suppression case (the real leak that must still fire).
What you should do
- Run a free scan of the production URL. Bundle secrets, client env names, system-prompt leaks, and a vulnerable RSC pin show up there.
- If the app has auth-gated routes, an MCP endpoint, GraphQL, or cloud storage URLs in the client, run a deep scan. That is where the read-only probes live. Pricing is on /pricing.
- Treat a hard-block secret as already public: rotate it, then remove it from the bundle. Deleting the line does not unpublish a key the browser already downloaded.
Frequently asked questions
Does the November 2026 Sekrd update send exploit payloads?
No. Bundle checks only read JavaScript the site already served. Deep-scan probes are read-only: they do not call MCP tools, do not send a React2Shell payload, and do not report a Next.js middleware bypass unless a path that was behind login actually returns the application.
Will Sekrd flag my Google Maps or Firebase API key?
A bare AIza key stays a low, tentative note because Maps, Firebase, and reCAPTCHA keys are meant to be in the browser. The critical finding fires only when that key is used with generativelanguage.googleapis.com, which is the billed Gemini API.
Is a system prompt in the JavaScript bundle always a vulnerability?
No. OWASP LLM07:2025 says the prompt itself is not a secret. Sekrd reports it only when the prompt text contains credentials or an authorization rule, such as an instruction to bypass Row Level Security.
Which new checks run on a free scan?
Provider key formats, sb_secret_, a Gemini key on the paid API, client-prefixed server secrets, system prompts that contain secrets or authorization logic, and a vulnerable React Server Components version pin. The middleware, MCP, GraphQL, bucket, .env, and .git probes run on a deep scan.
Sources
Don't ship until you're sekrd
Run a free scan to find the vulnerabilities your AI missed.
Scan Your App Free