# Cynbit Technologies: Backend Technical Audit Backend Developer Intern assessment, Phase 2. Passive analysis of www.cynbit.in, carried out on 1 November 2026. ## Executive summary Cynbit's public site is a server-rendered Next.js (App Router) application backed by Prisma and a relational database, with transactional email sent through Resend. It exposes at least six public, unauthenticated write surfaces: the contact form, the product demo form, a 25-field job application form, two newsletter sign-ups, a student-diary submission flow and blog comments. The most important observations are about **data handling**, not code. The careers form collects date of birth, gender, permanent address, phone and WhatsApp number as mandatory fields, while the privacy policy states that Cynbit "only collect[s] the minimum information necessary". That gap is both a compliance exposure under India's DPDP Act 2023 and the largest breach-impact multiplier on the site. The second theme is **abuse resistance**. Every write surface is public and none shows a visible bot challenge in the rendered markup. Combined with an email provider in the loop, that means spam, inbox flooding and quota exhaustion are the realistic threats, not sophisticated intrusion. The third theme is **server-side rendering hygiene**. Metadata is generated inconsistently across routes (duplicated title suffixes, every page claiming the homepage as its URL, job pages with no metadata at all), which points to the metadata API being used unevenly, and to some routes being client components when they could be server components. Twenty findings follow, each with evidence, impact and a concrete fix. Reference implementations for every fix live in [`recommendations/`](recommendations/). | Severity | Count | | --- | --- | | Critical (conditional) | 1 | | High | 1 | | Medium | 10 | | Low | 6 | | Info | 2 | ## Scope and method Only passive, authorised observation was used, in line with the brief. Specifically: - Eight public pages were loaded with ordinary GET requests, the same as a browser visit: `/careers`, `/careers/backend-developer-intern`, `/contact`, `/products`, `/privacy`, `/diaries`, `/blog/future-saas-development`, plus link structure from the navigation. - Public statements made by Cynbit itself (privacy policy, product pages, job descriptions) were used as architecture evidence. - No form was submitted, no endpoint was called directly, no paths were enumerated, no payloads were fuzzed and no authentication was attempted. **Limitation.** The fetch tooling used for the first pass returns rendered page content but not raw response headers or network payloads. Findings that depend on headers or on what the server does after a form submit are therefore marked **To verify**, and the repository ships a polite, read-only script ([`scripts/passive-audit.mjs`](scripts/passive-audit.mjs)) plus a DevTools checklist ([`evidence/README.md`](evidence/README.md)) that closes those gaps without sending anything a normal visitor would not. **Confidence labels used below** - **Observed**: directly visible in the public site as of the audit date. - **Inferred**: a reasonable conclusion from observed behaviour or Cynbit's own disclosures, not proven. - **To verify**: a standard control that cannot be seen from content alone; check with the script or internally. **Severity** combines likely impact with how easily it could happen. "Critical (conditional)" means critical only if the stated condition holds. ## Observed architecture ``` ┌──────────────────────────────┐ Browser ──HTTPS──────▶ │ www.cynbit.in (Next.js) │ │ App Router, SSR / SSG │ │ │ GET pages ──────────▶ │ RSC render + metadata API │ /_next/image ───────▶ │ Image optimiser ──────────────▶ images.unsplash.com │ │ POST forms ─────────▶ │ Server Actions or │ contact, demo, │ Route Handlers (unknown) │ careers (25 fields), │ │ │ │ newsletter x2, │ ▼ ▼ │ diary, comments │ Prisma ORM Resend API ───▶ inboxes └────────│─────────────────────┘ ▼ Relational DB (likely Postgres) ``` | Layer | Evidence | Confidence | | --- | --- | --- | | Next.js | `next-size-adjust` meta tag emitted by `next/font` on every page; `/_next/image?url=…` image URLs | Observed | | App Router | Metadata API title template behaviour (`%s \| Cynbit Technologies`) applied per route | Inferred | | Server rendering | Full page copy present in the initial HTML response | Observed | | Prisma + database | Privacy policy, section 4: "Prisma for database management" | Observed (self-disclosed) | | Resend email | Privacy policy, section 4: "Resend for email delivery" | Observed (self-disclosed) | | Stored user content | Privacy policy, section 3: "blog comments or diary entries" | Observed (self-disclosed) | | Remote images | Blog hero proxied from `images.unsplash.com` through `/_next/image` | Observed | | Hosting | Not determinable from content; check `server` and `x-vercel-*` headers | To verify | ### Public write surfaces | Surface | Location | Fields observed | Notes | | --- | --- | --- | --- | | Project inquiry | `/contact` | name, email, company, industry, message | Promises a 24-hour response | | Demo request | `/products#inquiry` | name, work email, company, message, product selector | Three products share one form | | Job application | `/careers/[slug]#application-form` | ~25 fields incl. DOB, gender, permanent address, phone, WhatsApp, repeatable qualifications | Highest-value PII on the site | | Newsletter | Site footer and blog posts | email | Two separate entry points | | Student diary | `/diaries` ("Share my entry") | Not visible; links to `/bitelevate` | Stored and published content | | Blog comments | Mentioned in privacy policy | Not visible on sampled post | Stored user content | ## What is already working well - All traffic observed was served over HTTPS on a single canonical host (`www`). - Pages are server-rendered, so content is in the first response. That is good for crawlers, link previews and time-to-content. - Résumés are collected as links rather than file uploads, which removes an entire class of upload vulnerabilities (malware, polyglot files, storage cost). - Remote images go through the Next.js optimiser instead of being hot-linked at full size. - The privacy policy names its processors and offers access and deletion rights, which is more than many small companies do. - Each role has its own stable slug (`/careers/backend-developer-intern`), a good basis for a proper jobs data model. ## Findings index | ID | Area | Finding | Severity | Confidence | | --- | --- | --- | --- | --- | | S-01 | Data handling | Careers form collects more personal data than the privacy policy describes | High | Observed | | S-02 | Data handling | Consent captures Terms only; no retention period or grievance contact | Medium | Observed | | S-03 | Security | Six public write surfaces with no visible bot protection | Medium | Observed | | S-04 | Security | Email-sending forms can be abused to flood third-party inboxes | Medium | Inferred | | S-05 | Security | Framework patch level for the December 2025 RSC and 2025 middleware advisories | Critical (conditional) | To verify | | S-06 | Security | Response security header baseline | Medium | To verify | | S-07 | Security | Image optimiser remote patterns may be broader than needed | Low | To verify | | S-08 | Security | Stored user content and RSC payload over-exposure | Medium | Inferred | | S-09 | Security | Portfolio and résumé URLs accepted without visible constraints | Low | Inferred | | S-10 | Security | No published vulnerability disclosure route | Info | To verify | | A-01 | Validation | Pre-selected defaults and missing cross-field rules produce silently wrong data | Medium | Observed | | A-02 | API design | No visible error contract or duplicate-submit protection on a 25-field form | Medium | Inferred | | A-03 | Architecture | Email delivery likely coupled to the request path | Medium | Inferred | | A-04 | Architecture | Job postings and applications need a relational model | Low | Observed | | A-05 | Architecture | Multi-tenant isolation for TenantEase and Revio | Info | Opportunity | | P-01 | Rendering | Metadata generated inconsistently across routes | Medium | Observed | | P-02 | Rendering | Social preview image is an SVG | Low | Observed | | P-03 | Performance | Metadata points at the apex host while the site serves `www` | Low | Observed | | P-04 | Performance | Icon ligature font leaks text and adds weight | Low | Observed | | P-05 | Performance | Caching strategy for query-string blog filters | Medium | To verify | ## Security and data-handling findings ### S-01. Careers form collects more personal data than the privacy policy describes **Severity:** High | **Confidence:** Observed **Evidence.** `/careers/backend-developer-intern` marks these as required: phone, WhatsApp number, date of birth, gender, permanent address, current city, LinkedIn URL and portfolio URL, plus full academic history. The privacy policy (section 1, updated 20 May 2026) says Cynbit collects "typically your name, email, and professional context". **Why it matters.** Date of birth, gender and permanent address are not needed to screen an intern application. Under the DPDP Act 2023, personal data must be limited to what is necessary for the stated purpose, and the notice must match what is collected. Practically, every unnecessary field enlarges the damage of a database leak, a misconfigured admin export or a compromised Resend notification. **Recommendation.** 1. Remove date of birth, gender and permanent address from the application stage. Collect them later, at offer stage, only if payroll or compliance needs them. 2. Merge phone and WhatsApp into one field with a "this number is on WhatsApp" checkbox. 3. Encrypt remaining contact PII at the application layer and store a keyed hash for lookups ([`08-pii-encryption/field-encryption.ts`](recommendations/08-pii-encryption/field-encryption.ts)). 4. Update the privacy notice so it lists exactly what the careers form collects. The minimised schema is in [`02-validation/application.schema.ts`](recommendations/02-validation/application.schema.ts). ### S-02. Consent captures Terms only; no retention period or grievance contact **Severity:** Medium | **Confidence:** Observed **Evidence.** The application's consent checkbox links to `/terms` but not to `/privacy`. The blog newsletter says "By subscribing, you agree to our Privacy Policy" without linking it. The privacy policy states no retention period and names no grievance officer; requests go to the general `opportunities@` inbox. **Why it matters.** Consent is only meaningful if the person can see what they are agreeing to, and the server needs proof of what version they agreed to and when. Without a retention rule, applicant data accumulates indefinitely. **Recommendation.** Store a consent record with each submission (policy version, timestamp, purpose, hashed IP), link the privacy notice from every form, publish a retention period, and enforce it with a scheduled purge ([`11-retention/purge-expired.ts`](recommendations/11-retention/purge-expired.ts)). The `Consent` model is in [`07-data-model/schema.prisma`](recommendations/07-data-model/schema.prisma). ### S-03. Six public write surfaces with no visible bot protection **Severity:** Medium | **Confidence:** Observed (surfaces), To verify (server-side controls) **Evidence.** Contact, demo request, job application, two newsletter forms and diary submission are all reachable without an account. None renders a visible CAPTCHA or challenge widget in the page content. Server-side rate limiting may exist but cannot be seen. **Why it matters.** Unauthenticated writes that land in a database and trigger email are the classic target for spam bots. Effects: polluted lead and applicant data, staff time lost to triage, database growth and, most expensively, email-provider reputation and quota (see S-04). **Recommendation.** Layer cheap controls, cheapest first: 1. Honeypot field plus a minimum form-fill time check (blocks naive bots at zero UX cost). 2. Per-IP and per-email sliding-window rate limits at the edge ([`05-abuse-protection/rate-limit.ts`](recommendations/05-abuse-protection/rate-limit.ts)). 3. Invisible Cloudflare Turnstile, verified server-side ([`05-abuse-protection/turnstile.ts`](recommendations/05-abuse-protection/turnstile.ts)). 4. A request body size cap so a 25-field form cannot be used to post megabytes. All four are wired together in [`03-api/applications.route.ts`](recommendations/03-api/applications.route.ts). ### S-04. Email-sending forms can be abused to flood third-party inboxes **Severity:** Medium | **Confidence:** Inferred **Evidence.** Resend is a disclosed processor, and the contact page promises a "24 H Response Sync". If any form sends an automatic acknowledgement to the submitted address, an attacker can type a victim's address and make Cynbit's domain send them mail repeatedly. **Why it matters.** This turns Cynbit into an email-bombing relay. The cost lands on Cynbit: spam complaints, damaged sender reputation on `cynbit.in` (affecting real outreach), and suspension risk at the email provider. **Recommendation.** - Send acknowledgements only after a bot check passes, and cap them per recipient per day. - Never echo free-text user input into an outbound email to an unverified address. - For the newsletter, use double opt-in: one confirmation email, then nothing until the link is clicked ([`05-abuse-protection/double-opt-in.ts`](recommendations/05-abuse-protection/double-opt-in.ts)). ### S-05. Framework patch level for the December 2025 RSC and 2025 middleware advisories **Severity:** Critical (if unpatched) | **Confidence:** To verify **Evidence.** The site runs Next.js with React Server Components. The exact version is not determinable passively, and was deliberately not fingerprinted or probed. **Why it matters.** CVE-2025-55182 ("React2Shell"), disclosed 3 December 2025, is an unauthenticated remote code execution flaw in the React Server Components protocol; Next.js tracked its impact as CVE-2025-66478 for App Router apps on 15.x and 16.x. Separately, CVE-2025-29927 (March 2025) allowed requests to skip Next.js middleware, which matters if any admin or dashboard route relies on middleware alone for authorisation. **Recommendation.** - Confirm `next` is on a patched release for its line (for example 15.5.7+ or 16.0.7+, or newer patch releases issued since) and that `react-server-dom-*` is 19.0.1, 19.1.2, 19.2.1 or later. - Enable Dependabot or Renovate with security updates auto-merged after CI. - Never rely on middleware as the only authorisation check; re-check the session inside each route handler or server action ([`10-multi-tenancy/tenant-scoped-prisma.ts`](recommendations/10-multi-tenancy/tenant-scoped-prisma.ts) shows the pattern). ### S-06. Response security header baseline **Severity:** Medium | **Confidence:** To verify **Evidence.** Headers were not captured by the first-pass tooling. Next.js does not set CSP, HSTS, `Permissions-Policy` or frame protection by default, so unless they were added deliberately they are likely absent. **Why it matters.** A Content-Security-Policy is the main safety net if stored content (S-08) ever renders unescaped. HSTS prevents downgrade on first visit from hostile networks. `frame-ancestors` stops the job application form being framed for clickjacking. **Recommendation.** Run `npm run audit` (see the evidence checklist), then apply the baseline in [`01-security-headers/next.config.ts`](recommendations/01-security-headers/next.config.ts): ```ts { key: "Strict-Transport-Security", value: "max-age=63072000; includeSubDomains; preload" }, { key: "X-Content-Type-Options", value: "nosniff" }, { key: "Referrer-Policy", value: "strict-origin-when-cross-origin" }, { key: "Permissions-Policy", value: "camera=(), microphone=(), geolocation=()" }, { key: "Content-Security-Policy", value: csp }, // frame-ancestors 'none', object-src 'none', ... ``` Roll out CSP as `Content-Security-Policy-Report-Only` first, watch reports for a week, then enforce. ### S-07. Image optimiser remote patterns may be broader than needed **Severity:** Low | **Confidence:** To verify **Evidence.** Blog images are served as `/_next/image?url=https://images.unsplash.com/…&w=1920`. **Why it matters.** If `images.remotePatterns` uses a wildcard host, `/_next/image` becomes a free image-resizing proxy for any URL on the internet, billed to Cynbit's hosting account and usable to amplify load. **Recommendation.** Pin the allowed host and path, and restrict output sizes: ```ts images: { remotePatterns: [{ protocol: "https", hostname: "images.unsplash.com", pathname: "/photo-**" }], deviceSizes: [640, 960, 1280, 1920], } ``` Longer term, self-host hero images so a third-party CDN is not on the critical rendering path. ### S-08. Stored user content and RSC payload over-exposure **Severity:** Medium | **Confidence:** Inferred **Evidence.** Diaries and blog comments are stored and published (privacy policy, section 3). Diary entries display as "Anonymous Student". **Why it matters.** Two distinct risks: 1. **Stored XSS.** If entries are rendered through `dangerouslySetInnerHTML` or a permissive Markdown renderer, a single malicious entry runs script for every visitor. 2. **Hidden data leakage.** In the App Router, any prop passed to a client component is serialised into the page's RSC payload (`self.__next_f`). If a full Prisma row is passed down, the submitter's email or name ships to every browser even though the UI shows "Anonymous". It is invisible in the rendered page and visible in view-source. **Recommendation.** Moderate before publish (status `PENDING → APPROVED`), render as plain text or with an allow-list sanitiser, and always map database rows to a public DTO with explicit `select`: ```ts const entries = await db.diaryEntry.findMany({ where: { status: "APPROVED" }, select: { slug: true, title: true, excerpt: true, persona: true, publishedAt: true }, // never: authorEmail, authorName, ipHash }); ``` ### S-09. Portfolio and résumé URLs accepted without visible constraints **Severity:** Low | **Confidence:** Inferred **Evidence.** The application requires "Portfolio / Resume Link" and "LinkedIn Profile URL" as free-text fields. **Why it matters.** Today the risk is data quality (non-URLs, `javascript:` links that become clickable in an admin view). If a future feature fetches these links server-side for previews or AI résumé parsing (Cynbit Studio advertises an "AI Resume Engine"), unvalidated URLs become a server-side request forgery vector into internal or cloud-metadata addresses. **Recommendation.** Accept `https:` only, validate LinkedIn against `linkedin.com/in/…`, render admin links with `rel="noopener noreferrer"`, and if links are ever fetched, resolve DNS and refuse private, loopback and link-local ranges. The URL rules are in the validation schema. ### S-10. No published vulnerability disclosure route **Severity:** Info | **Confidence:** To verify **Recommendation.** Publish `/.well-known/security.txt` (RFC 9116) with a dedicated `security@` contact so researchers have a safe, private route instead of public posts. Template: [`12-disclosure/security.txt`](recommendations/12-disclosure/security.txt). ## API, validation and architecture findings ### A-01. Pre-selected defaults and missing cross-field rules produce silently wrong data **Severity:** Medium | **Confidence:** Observed (defaults as rendered; confirm visually) **Evidence.** In the rendered application form, the dropdowns appear to arrive with values already chosen: Bachelor's Degree, B.Tech / B.E., graduation year 2025, education status Completed, score system CGPA. The graduation-year list runs to 2030, so "Completed" plus "2029" is selectable. Score Value is a single field whose valid range depends on whether CGPA (0 to 10) or percentage (0 to 100) is picked. **Why it matters.** A required field with a default is not really required: applicants who skip it submit wrong data that looks valid. Recruiters then filter on bad values. Without cross-field rules, the database accepts impossible combinations. **Recommendation.** Use an empty placeholder option for every required select, and share one schema between client and server so the same rules run in both places: ```ts .superRefine((q, ctx) => { const max = q.scoreSystem === "CGPA" ? 10 : 100; if (q.scoreValue > max) ctx.addIssue({ code: "custom", path: ["scoreValue"], message: `Must be at most ${max} for ${q.scoreSystem}` }); if (q.status === "COMPLETED" && q.graduationYear > CURRENT_YEAR) ctx.addIssue({ code: "custom", path: ["graduationYear"], message: "A completed qualification can't end in the future" }); }) ``` Full schema: [`02-validation/application.schema.ts`](recommendations/02-validation/application.schema.ts). ### A-02. No visible error contract or duplicate-submit protection on a 25-field form **Severity:** Medium | **Confidence:** Inferred **Why it matters.** Long forms fail in specific ways: a slow mobile connection leads to a double tap on "Submit Application Blueprint", producing two applications; a validation failure that returns a generic error makes the applicant re-type 25 fields; a server error returns an HTML page the client cannot parse. **Recommendation.** - Return one machine-readable error shape everywhere, RFC 9457 `application/problem+json`, with per-field errors the form can map back onto inputs ([`04-errors/problem.ts`](recommendations/04-errors/problem.ts)). - Have the client generate an `Idempotency-Key` per submission attempt; the server stores it and replays the original response on retry. - Add a database unique constraint on `(jobPostingId, emailHash)` so the same person applying twice to one role updates rather than duplicates. - Version public endpoints (`/api/v1/…`) if they are Route Handlers, so the planned mobile app or partners can depend on them. Example error response: ```json { "type": "https://cynbit.in/problems/validation", "title": "Some fields need attention", "status": 422, "errors": [ { "path": "qualifications.0.scoreValue", "message": "Must be at most 10 for CGPA" } ], "requestId": "req_01JBX…" } ``` ### A-03. Email delivery likely coupled to the request path **Severity:** Medium | **Confidence:** Inferred **Why it matters.** The common pattern is "validate, insert row, call Resend, return 200". If Resend is slow, the user waits; if it errors after the insert, the lead exists but no one is notified; if the code sends before inserting, a database failure loses the lead entirely while the team gets an email for it. **Recommendation.** Transactional outbox: write the submission and an `OutboxMessage` row in one database transaction, return immediately, then dispatch from a background step with retries and an idempotency key sent to Resend so retries never double-send. Code: [`06-outbox/outbox.ts`](recommendations/06-outbox/outbox.ts) and [`06-outbox/dispatch.route.ts`](recommendations/06-outbox/dispatch.route.ts). ``` POST /applications ─▶ BEGIN; INSERT application; INSERT outbox; COMMIT ─▶ 201 │ cron / after() ─▶ SELECT … FOR UPDATE SKIP LOCKED ─▶ Resend (Idempotency-Key) └── on failure: attempts++, backoff ``` ### A-04. Job postings and applications need a relational model **Severity:** Low | **Confidence:** Observed **Evidence.** Twenty roles with distinct slugs; an "Add another qualification" control, which means one applicant has many qualifications; "Applications are reviewed within 48 hours", which implies a review workflow. **Recommendation.** Model postings in the database (or a CMS) rather than code so non-engineers can open and close roles, and render them with static generation plus on-demand revalidation. Store applications relationally: `JobPosting 1─N Application 1─N Qualification`, with a `status` enum (`RECEIVED, SCREENING, INTERVIEW, OFFER, REJECTED, WITHDRAWN`) and timestamps for the 48-hour SLA. Full Prisma schema: [`07-data-model/schema.prisma`](recommendations/07-data-model/schema.prisma). ### A-05. Multi-tenant isolation for TenantEase and Revio **Severity:** Info | **Confidence:** Opportunity (not observable from the marketing site) **Context.** The product pages describe TenantEase storing tenant ID verification and rent transactions, and Revio offering Admin, Manager and Advisor roles across locations. Both are multi-tenant SaaS, where the costliest bug class is one customer seeing another's data. **Recommendation.** Enforce tenancy in two independent layers: an application-layer Prisma extension that injects `organizationId` into every query, and Postgres row-level security as a backstop if application code is ever wrong. Check role permissions on the server for every mutation, never only in the UI. Store ID documents in private object storage, served through short-lived signed URLs. Code: [`10-multi-tenancy/`](recommendations/10-multi-tenancy/). ## Rendering and performance findings ### P-01. Metadata generated inconsistently across routes **Severity:** Medium | **Confidence:** Observed **Evidence.** | Route | Problem | | --- | --- | | `/careers` | Title "Careers \| Cynbit Technologies \| Cynbit Technologies" (suffix applied twice) | | `/contact`, `/products`, `/privacy` | Same doubled suffix | | All sampled static routes | `og:url` is `https://cynbit.in` regardless of the page | | `/careers/backend-developer-intern` | No description, no Open Graph tags; title appears at the end of `
` instead of in `` | | `/diaries` | Falls back to the homepage title and description | | Blog posts | No `og:url` at all | **Why it matters.** Sharing any job link on LinkedIn or WhatsApp, which is how internships spread, shows no description and points the preview at the homepage. Search engines see many pages declaring the same URL. A `