Biztonságos alkalmazásfejlesztés

Secure Coding

Threat model first

Spend five minutes thinking like an attacker before adding controls:

• Trust boundaries: where untrusted data enters (requests, forms, uploads, webhooks, third-party APIs, queues, model output, files or paths from other processes). Trust follows who wrote a value, not the channel that delivered it.
• Assets: credentials, personal data, payment data, admin actions, money movement.
• For each boundary ask: can someone impersonate a user or service, alter data, deny an action later, read data they should not, overwhelm it, or gain rights? Name the mitigation for each.
• Write the abuse case next to the use case and make it a test.

Ask a human first

New or changed authentication flows, storing a new kind of sensitive data, a new external integration, CORS changes, new upload handlers, changes to rate limits, granting elevated roles.

Secrets

• Never put secrets, API keys, tokens, passwords or private keys in source, tests, fixtures, logs, commit messages or client bundles.
• Read them from the environment or the repo's secret mechanism; add placeholders, never real values, to the example env file.
• Do not log tokens, cookies, auth headers, full request bodies or personal data. Redact.
• If you find a committed secret, do not repeat its value; report it so it can be rotated.

Input and output

• Treat all external data as untrusted: params, bodies, headers, cookies, files, webhooks, third-party API responses.
• Validate on the server with allowlists: type, length, format, range, enum. Client-side validation is only UX.
• Encode output for its context (HTML, attribute, URL, JS). Prefer the framework's auto-escaping; justify every raw render.
• Never build shell commands, eval, regexes or templates from user input. Pass process arguments as arrays.
• Deserialize only safe formats such as JSON; never unserialize or unpickle untrusted data.

Database

• Use parameterized queries or the ORM/query builder bindings. Never concatenate input into SQL.
• Identifiers (columns, sort direction, table names) cannot be bound: map them through an allowlist.

Authentication and authorization

• Check authorization server-side on every endpoint, action and query, not only in the UI.
• Verify the current user may access the specific record (prevent IDOR); scope queries by user or tenant.
• Deny by default: new routes must not become public by accident.
• Use the framework's sessions, password hashing (bcrypt, argon2) and CSRF protection. Never roll your own crypto.
• Session cookies: HttpOnly, Secure, SameSite.

Files and URLs

• Never join user input into file paths without normalizing and confirming the result stays inside the allowed base directory.
• Uploads: enforce size limits, check type by content not only extension, generate the stored name, store outside the web root or without execute permission.
• When the server fetches user-supplied URLs, allowlist hosts and block internal addresses (SSRF).

Dependencies and privileges

• Add dependencies only when needed; prefer maintained, widely used packages, locked through the lockfile.
• Do not disable TLS verification, CSP, security headers or CORS restrictions to make something work. Never reflect any origin with credentials.
• Least privilege: minimal API scopes, roles, file permissions and database grants.

Errors

• Show users generic messages; log details server-side. No stack traces, SQL or internal paths in responses.
• Fail closed: if an auth or validation check errors, deny.

Before you finish

• No secrets or personal data in code, logs, tests or commits.
• New input validated server-side; output encoded for its context.
• Queries parameterized; dynamic identifiers allowlisted.
• Every new or changed endpoint checks authentication and record-level permission.
• File paths, uploads and outbound URLs constrained.
• Error responses leak nothing internal.