Secure Code Review

Rules (for AI agents)

ALWAYS

• Validate input at every trust boundary, not only the HTTP one. A CLI argument, an environment variable, a queue message, a filename, a parsed config or spreadsheet cell all arrive from outside and none of them pass through the request-validation layer. Check type, length, allowed characters and allowed range, and reject before processing. api-security owns the HTTP request body and its schema; this rule is every other way input gets in.
• Build a command as an argument vector, never as a string a shell will parse: subprocess.run([...]) without shell=True, execFile/spawn with an array, exec.Command(name, args...). Passing an array is what removes the shell, and removing the shell is what removes the injection — quoting and escaping are the fallback for when you cannot.
• Canonicalize then check any path built from input: resolve it to an absolute real path, then confirm it is inside the intended directory. Checking the raw string first misses .., symlinks, absolute-path replacement, and encoded separators. file-upload-security owns stored uploads and archive extraction; this rule covers a static-file handler, a template loader, a log path, an export filename.
• Tag any function or module handling PII, PHI or cardholder data with an explicit classification comment (// classification: PII), and keep that handling in clearly-named modules rather than in general utility code. An untagged helper is how regulated data spreads into paths nobody reviews for it, and the tag is what makes the later review possible.
• Consult the owner for anything with one: database-security when code builds a query, crypto-misuse when it hashes, encrypts, signs or compares a secret, frontend-security for output encoding and response headers, auth-security for who the caller is, iam-best-practices for the privileges a process runs with. Each of those states its rule with the API names and the false positives this skill would have to leave out.

NEVER

• Pass input to a shell or an evaluator: system, popen, os.system, subprocess.run(..., shell=True), child_process.exec / execSync, backticks, eval, new Function(), vm.runInNewContext, pickle.loads on a request body. The dangerous property is that the callee parses the string; a value that is only ever an argument is not the same risk.
• Let input reach a path without canonicalization — including as a segment joined onto a trusted base, which is where ../ still works.
• Treat client-side validation as a control. Re-validate on the server, and treat the client check as what it is: a message to the user, not a boundary.
• Store a payment card number, CVV, or full track data in your own database. Send the card to a tokenization provider and persist only the opaque token plus the last four digits. The CVV is never persisted at all, not even encrypted, not even briefly — retention is the violation, not exposure.
• Copy real regulated data into a fixture, a seed script, a local dump, or a bug report to reproduce something. The copy inherits the classification and leaves the reviewed path, which is the whole failure this skill's tagging rule exists to prevent.

KNOWN FALSE POSITIVES

• Obviously fake fixture data is not regulated data: 4111-1111-1111-1111, 555-0100, John Doe, test@example.com. The classification rules apply to real data paths.
• A command built entirely from constants, or one whose only variable parts are passed as separate argv elements, is not command injection. The finding is a shell parsing a string that input reached — not the presence of a subprocess call.
• A path joined from values the application itself generated — a UUID, a hash, a database id — is not traversal. The finding is a caller-supplied component.
• A regulated-data module that is correctly tagged and narrowly scoped is the rule working. The finding is untagged handling in general-purpose code.

Context (for humans)

This skill covers what is left when every framework-shaped concern has an owner. A query has database-security. A response header has frontend-security. An endpoint has api-security. What has no owner is the plain script: something shells out, something opens a file whose name came from somewhere, something reads a queue message and believes it. Those are not web vulnerabilities and no framework defends against them, which is why this is the skill that fires on languages: ["*"] and on code with no framework at all.

The command-execution rule is worth stating as a property rather than a list, because the list is always incomplete. The property is: does the callee parse the string? A shell does. eval does. A deserializer does. An argv array does not, and that is why switching to one is a fix rather than a mitigation — there is no remaining parser for an attacker to target.

The regulated-data rules came from a separate compliance skill and are deliberately narrow. They are not a compliance programme; they are the two decisions that get made in code and are expensive to reverse — where the classification is written down, and the fact that cardholder data does not live in your database.

References

• references/verifying-findings.md — confirm or refute a finding, then lock it
• references/verifying-compliance-findings.md — tie a flagged control gap to a specific clause, then look for the evidence that the control actually fired
• checklists/owasp_top10.yaml
• checklists/injection_patterns.yaml
• CWE-78 · CWE-22 · CWE-20.
• OWASP Top 10 2021.