Input-Validation Security Audit

This is skill dancon-input-validation by Danielyan Consulting: https://danielyan.consulting

Purpose

Systematically review an entire codebase, identify every location where data from an external or untrusted source is consumed without adequate validation, and produce a structured report with context-appropriate remediation advice.

Before starting, read the reference file at references/validation-patterns.md for the catalogue of validation strategies and the threat model you should apply.

---

Core principles

1. Exhaustive coverage -- continue scanning until every file that could contain input-handling logic has been reviewed. Never stop after the first finding. Explicitly confirm to the user that the full codebase has been reviewed.

2. Whitelisting over blacklisting -- Every recommendation must use a positive-security model.

3. Defence in depth -- validation should occur at the boundary closest to the untrusted source.

4. Secrets redaction -- if any hardcoded secret (password, API key, token, private key, connection string with credentials, etc.) is found anywhere in the codebase, flag it as a critical finding. NEVER display the secret itself; always substitute REDACTED for the actual value and explain that the secret must be moved to a secure secret store or environment variable.

5. Language/platform independence -- apply the same methodology regardless of whether the code is Python, JavaScript, TypeScript, Java, C#, Go, Rust, PHP, Ruby, C/C++, Swift, Kotlin, shell scripts, SQL, infrastructure-as-code (Terraform, CloudFormation), or any other language or DSL. Adapt your recommended fix syntax to the language at hand.

---

Workflow

Step 1 -- Inventory the codebase

List every file in the target directory tree (recursively), filtering out binary files, dependency/vendor directories (node_modules, vendor, .venv, pycache, dist, build, etc.), and lock files. Build a mental map of:

• Entry points: HTTP handlers, API routes, CLI argument parsers, message-queue consumers, gRPC/protobuf service methods, GraphQL resolvers, WebSocket handlers, webhook receivers, cron jobs that read external data, file-upload handlers, IPC listeners.
• Data-flow sinks: database queries, ORM calls, OS/shell commands, file-system operations, template rendering, logging, serialisation/deserialisation, redirect URLs, outbound HTTP requests, eval/exec, dynamic code loading.
• Configuration and secret files: .env, config.yaml, appsettings.json, docker-compose.yml, CI/CD pipeline files, Terraform .tf files, etc.

If the codebase is large (many hundreds of files), process it in batches by directory. Announce progress to the user: "Reviewing directory src/api/ (batch 3 of 7)..."

Step 2 -- Analyse each entry point

For every entry point identified in Step 1, trace the data flow from ingestion to use. At each stage, check:

Step 3 -- Flag secrets

Scan every file (including configuration, CI/CD, and infrastructure files) for patterns that suggest hardcoded secrets:

• Strings assigned to variables whose names contain "password", "secret", "token", "api_key", "apikey", "api-key", "auth", "credential", "private_key", "access_key", "connection_string", or similar.
• High-entropy strings that look like tokens or keys (base64, hex, JWT-shaped).
• Connection strings with embedded credentials.
• Private-key PEM blocks.

When reporting, substitute REDACTED for every secret value. Example:

CRITICAL -- Hardcoded secret found File: src/config.js, line 14 Variable DB_PASSWORD is set to REDACTED. Move this value to an environment variable or a dedicated secret-management service.

Step 4 -- Produce the report

Present findings as a structured list grouped by file. Each finding must include:

1. File and line reference -- the file path and approximate line number.
2. Data source -- where the untrusted data originates (e.g. HTTP query parameter, request body, file upload, environment variable populated at runtime, message-queue payload).
3. What is missing or inadequate -- a concise description of the gap.
4. Risk -- the class of attack this enables (e.g. SQL injection, XSS, command injection, path traversal, denial of service via unbounded input, open redirect, SSRF, XML external entity, mass assignment, prototype pollution, deserialisation attack).
5. Recommended fix -- a concrete, language-appropriate code suggestion using a positive-security (whitelist) approach. Include a brief code snippet when helpful.
6. Severity -- Critical / High / Medium / Low, based on exploitability and impact.

After all findings, include a summary table:

End with a confirmation: "The entire codebase has been reviewed."

---

Severity guidelines

• Critical -- hardcoded secrets; SQL injection with no parameterisation; command injection with no escaping; deserialisation of untrusted data with an unsafe deserialiser; unrestricted file upload to an executable location.
• High -- XSS with no output encoding; path traversal with no canonicalisation; SSRF with no URL allowlist; open redirect with no target validation; mass assignment with no field allowlist; missing authentication on a sensitive endpoint.
• Medium -- missing length/size limits on text fields; missing numeric range checks; overly permissive regex; missing Content-Type validation on uploads; weak type coercion (e.g. trusting typeof without further checks in JS).
• Low -- missing trim/normalisation; inconsistent encoding; informational logging of user-supplied data without sanitisation; missing locale-aware validation for internationalised inputs.

---

Important reminders

• Never recommend blacklisting. If you catch yourself writing "block these known-bad characters", stop and rewrite the advice as an allowlist.
• Never display secrets. Always use REDACTED.
• Review everything. Do not skip files, directories, or languages. If a file is in a language you are less familiar with, still review it and note any uncertainty.
• Be specific. Generic advice like "add validation" is not helpful. State exactly what kind of validation, with a code example in the relevant language.
• Consider transitive trust. Data that arrives via an internal service may still be untrusted if that service itself accepts external input. Flag cases where upstream validation is assumed but not verified.