En

Feature-scoped security review (single-feature security audit)

For the-platform project: a microservices CRM platform that processes customer personal data. Security is a critical priority, not a formality. The audit must find real, exploitable problems tied to file:line, not produce a generic checklist without verification. Every finding must be confirmed by hand, not merely mentioned in a scanner's output.

This is the focused version of the full repository audit (see the security-audit-full skill if the task is the whole repository rather than a single feature). The verification principles are the same, but the scope, findings, and report are strictly limited to the code that belongs to this feature and to what it touches. The manual-analysis work can be delegated through the Agent tool — use parallelization by zone as described in the "Running the check" section below.

INPUT: HOW TO DETERMINE THE FEATURE

Feature: $ARGUMENTS

The feature is passed in one of three forms — figure out which one you have and build the review scope accordingly. The scope is ALWAYS wider than the literal input: include direct consumers/calling code (the router that registers the handler; the frontend that hits the API; the neighboring service that receives the inter-service call).

A. DIRECTORY/BRANCH/DIFF (e.g. services/xxx-service/feature_y/ or "diff between dev and the feature/PROJ-XXXX branch"):

• Scope = the entire contents of the directory (or the files from git diff --stat against the base branch) + modules that import it (grep -r on the package/module name outside the directory) + the routes/DI that register it (main.py/app factory/router include).
• If the directory is a shared library (libs/shared_auth, libs/shared_metrics, etc.), be sure to identify ALL consumers of the library across all services — a vulnerability in shared code multiplies across the whole list of consumers.

B. DOCUMENT (path to a spec/design document/PRD, .md/.txt/.docx):

• Read the document in full. Extract from it: endpoint/route names, model/table names, roles and permissions, UI component/screen names, and any mentioned external integrations (webhooks, callback URLs, third-party APIs).
• For each extracted term, grep/search the codebase to translate the "what should exist" description into concrete file:line "what actually exists". Do not settle for the document saying something is "implemented" — verify the code, not the document text.
• If the document describes intent rather than fact (a draft spec), flag explicitly in the report which items have no counterpart in the code (this is a finding too: a mismatch between spec and implementation may mean an unfinished access control).

C. YOUTRACK ISSUE (an ID of the form PROJ-XXXX or a link):

• Fetch the issue text (title, description, comments, acceptance criteria) through whatever tracker-integration mechanism the project provides (a YouTrack/Jira/GitHub/Linear MCP tool, if connected). If no programmatic access exists, ask the user directly for the issue text and links to related PRs — do not invent the content.
• Find related commits and files by ticket ID: commits in this repository are conventionally tagged with the ticket number in the message (e.g. PROJ-1042, PROJ-1031, PROJ-318) — use git log --all --grep=<ISSUE-ID> --oneline, then git show --stat <hash> / git log --all -- <files from commit> to build the list of affected files and services.
• If the ticket references a PR/branch, review that branch's diff specifically (git diff main...<branch>), not just the final state of main, so you do not miss intermediate versions if the branch is not yet merged.

If none of the three sources determines the scope unambiguously, stop and explicitly list what needs to be clarified with the task author, rather than blindly reviewing the entire service.

Explicitly record at the top of the report the final scope (the list of files/directories/services) that resulted from this step — this is your working SCOPE, and the rest of the audit runs on it, plus the points of contact with the rest of the system (see PASS 3 below).

KEY PRINCIPLE: VERIFICATION, NOT TRUST

The reason feature checks fail is taking it on faith that "the code was written to spec" is equivalent to "the feature is secure". It is NOT. Verify adversarially:

1. If the feature adds a check (a flag, an HMAC signature, escaping, authorization), make sure it is ENABLED by default and active in all environments (dev/staging/prod), not merely implemented in code and turned off by a flag.
2. If the feature adds input escaping, verify COMPLETENESS (single quote, double quote, backslash, null byte, unicode bypass), not just the obvious case.
3. If the feature protects one endpoint/service, check whether it has "siblings" of the same class of operation elsewhere in the codebase that were left unprotected (e.g. the feature added auth on POST but not on DELETE of the same resource, or on an analogous bulk endpoint).
4. If the feature has RBAC/switch/if-else on roles, check the "else" branch (default case). The absence of an explicit access denial by default is a hole.
5. Do not accept the ticket/spec description ("this is already protected", "it uses the shared middleware here") without independently verifying the current code.
6. State the status explicitly: "not implemented" / "implemented on paper (code exists, no protection)" / "implemented selectively (part of the class covered)" / "fully implemented" / "new finding outside the feature's scope".

METHODOLOGY: THREE INDEPENDENT PASSES (within the feature SCOPE)

PASS 1 — Focused automated scanning

• If the feature added new dependencies (a new entry in requirements/pyproject/package.json), run pip-audit/npm audit focused on the changed lock file, not on the whole repository.
• semgrep/bandit on the files in SCOPE (injection, ssrf, insecure deserialization, hardcoded secrets, weak crypto).
• If the feature touches Dockerfile/helm-values/k8s manifests, run checkov/ kube-linter focused on the changed files.
• gitleaks/trufflehog on the feature's diff (git diff/git log -p on the affected files and commits) — a secret may have been committed and then deleted within the same branch.

PASS 2 — Manual line-by-line review of the feature code

Review every file in SCOPE line by line (not diagonally), applying the category checklist below — only those categories that actually apply to what the feature does (see the applicability hints before the checklist). For each finding: file:line, vulnerability type, a concrete exploitation scenario ("request X with parameter Y yields result Z"), severity, status.

If the SCOPE is large (several services/directories) and the Agent tool is available, split it into independent zones and use several subagents, one zone each, so you don't cut corners across the whole volume at once (see "Running the check" below).

PASS 3 — Points of contact with the rest of the system

Independently of the line-by-line review, answer: how does the feature fit into the platform's existing security invariants, not just "are there holes in its own code"?

• Does the feature use the existing authentication/authorization/multi- tenancy mechanisms, or does it invent its own path that bypasses them (a new handler that does not go through the shared auth middleware/RBAC decorator)?
• If the feature adds a new internal endpoint between services, is it protected by the same shared-secret/mTLS mechanism as the rest of the /internal/* endpoint class, or is it an exception?
• If the feature adds a new data read/write point, is it filtered by company_id/tenant_id the same way as the other points of the same type in the system?
• If the feature reuses a shared library (libs/shared_auth, etc.), does it use the current version as a single package, or did it copy/fork the logic into itself?
• Does the feature break any of the existing security invariants documented in previous repository audits? In this project that is not an abstract caveat — the registry of past findings lives in docs/bugs/security_audit/ (each past-audit finding records a verdict: confirmed / false positive / already fixed), and the regression check over it is scripts/verify_audit_fixes.py. If the feature SCOPE overlaps by topic or by files with one of the findings in that folder, be sure to run the script (if present in the repository) and explicitly reconcile the current status, rather than reopening the finding from scratch.
• If a feature finding effectively coincides with a risk already tracked in docs/bugs/security_audit/ that was deliberately accepted (a low replica count for HA, a disabled networkPolicy, etc.), reference the existing finding file instead of raising a duplicate as a "new feature finding"; note it only if the feature makes the risk worse than it was recorded.

CATEGORY CHECKLIST (apply the ones relevant to the feature)

Before the review, quickly classify the feature: which of the blocks below apply (a new API endpoint → blocks 1,2,5,6,10; a new UI screen → blocks 1,7; file upload → block 9; integration/webhook → blocks 4,6,8; infrastructure change → blocks 11,12). Do not skip a block just because "the feature is small" — small features are typically exactly the source of pinpoint holes (see "Edge cases" below).

1. Authentication and authorization

   • Does the new/changed endpoint require authentication the same way as the other endpoints of the same service (is there a "forgotten" open route)?
   • RBAC branches: what happens for a role that falls into no explicit branch (default/else)? Does this lead to a leak (no filter by assigned_user_id/owner_id)?
   • IDOR/BOLA: can you access someone else's object that the feature operates on by changing an ID in the URL/body?
   • Broken function level authorization: is the new administrative/bulk operation accessible without a role check?
   • Mass assignment/over-posting: does the new/changed API accept arbitrary request-body fields (role, is_admin, company_id, balance, etc.), or is the list of allowed fields an explicit whitelist?
   • Session/token: if the feature touches sessions/tokens — where is the storage, is there invalidation, is there a lifetime limit?
   • Rate limiting: if the feature adds a login/password reset/OTP-like flow — is there a limit on attempts?

2. Multi-tenancy (isolation between customer companies) — review this separately from generic IDOR; a cross-tenant leak is heavier in consequences.

   • Every new read/write point (REST handler, WS subscription, cache key, search index, queue, export) — is it filtered by company_id/tenant_id at the DB-query level, not just at the UI/router level?
   • Is the company_id match checked by an explicit comparison of the value from the token/context against the record's company_id, rather than assumed?
   • Shared resources (Redis keys, RabbitMQ queues, file storage) — can a key/name collide between companies given identical internal IDs?
   • A new Alembic migration/DDL: does a new table/column that stores a specific company's data have company_id/tenant_id and an index on it — a structural gap at the schema level cannot be caught by a line-by-line query review if the column itself is missing.

3. WebSocket / realtime (if the feature touches a realtime channel)

   • Is the token/role/company_id checked at connect AND separately at each channel/room subscription (join)?
   • Can a client subscribe to someone else's channel by supplying/guessing a room-id/dialog-id/user-id?
   • Is broadcast filtered by company_id/role before sending, or does the server rely on the client "just not being subscribed"?

4. Internal-API between services (if the feature adds/changes /internal/*)

   • Does it require a shared-secret/mTLS the same way as the other endpoints of the same class of operation in other services? Is the secret actually verified (not merely declared in env)?
   • Can the new internal endpoint be called directly from outside the cluster?

5. Injections (SQL / NoSQL / Command / Template / Deserialization)

   • Building queries by concatenation/f-string/format instead of parameterized queries/ORM — especially in new query parameters (filters, search, sorting).
   • "Manual" escaping — verify completeness (backslash, unicode, nested quotes).
   • Command injection: new subprocess/os.system/exec with user input.
   • Insecure deserialization of new data from a queue/webhook/inter-service call (pickle, yaml.load without SafeLoader, eval/exec).

6. SSRF (if the feature makes an outbound HTTP request to an external URL)

   • Is there a domain allow-list, blocking of private/internal IP ranges (169.254.x.x, 10.x, 172.16-31.x, 192.168.x, 127.x, cloud metadata)?
   • Is the webhook's signature/origin verified before the system initiates a response request to a URL from its body?

7. XSS (including client-side widgets, if the feature is a frontend/embeddable component)

   • innerHTML/dangerouslySetInnerHTML/document.write with untrusted data.
   • If the feature is a widget embedded on third-party sites: the vulnerability is visible to the end users of your platform's customers — qualify severity accounting for the expanded blast radius.
   • Are CSP/clickjacking headers not weakened by the new code?

8. Secrets and PII handling (if the feature touches integrations/tokens/PII)

   • New secrets in code/configs tracked by git.
   • New integration tokens in the DB — plaintext or encrypted?
   • Secret masking in new logs — does it work for the new data types the feature introduced?

9. File upload and storage (if the feature adds upload/download)

   • MIME-type/extension validation (do not trust the client's Content-Type).
   • Size limits, protection against zip bombs.
   • Path traversal when building a path from user input.
   • Public serving — does it require authentication if the files contain PII?

10. Dependencies and supply chain (if the feature added new packages)

    • Criticality of new dependencies, whether a safer alternative exists.
    • A new internal package — installed from a private index/local path, or by a bare name that can be substituted from public PyPI/npm (dependency confusion)?

11. Docker / containers (if the feature changes the Dockerfile)

    • Is USER set (non-root)? Are secrets not passed via ARG/ENV?

12. Kubernetes / Helm (if the feature changes values/manifests)

    • Are networkPolicy/securityContext/resources not weakened by the new values file relative to the project's existing baseline?
    • Secrets via Kubernetes Secrets, not plaintext in values.yaml?

13. Error handling and observability

    • A new broad except that logs and returns None/an empty result without re-raise — does this mask a failed authorization check as "allowed"?
    • Are new security events (failed authorizations, role changes) logged separately?
    • Leakage of internal information through new error messages to the outside.

14. CI/CD (if the feature changes the pipeline)

    • CI script injection: is unvalidated external input (PR title, branch name) interpolated directly into a run: step?
    • Does the new security gate actually block the merge, or does it only print a warning?

EDGE CASES OFTEN MISSED WHEN TESTING A FEATURE

• Functionality protected at the UI level (button hidden) but accessible directly through the API without a backend check.
• A race condition in "check-then-act" (the role check and the action are not atomic).
• Behavior for soft-deleted records — are they accessible through a new endpoint that does not account for the deletion flag?
• The bulk variant of the new feature (bulk endpoint) — often added "quick and dirty" with weaker-verified authorization than its single-item counterpart.
• The webhook the feature adds — is the signature/origin verified?
• The feature's feature flag, turned off "temporarily for debugging" and left off by default in the config.
• dev/staging/prod differences — a fix applied in one helm-values configuration may be absent in another.
• Data export, if the feature adds it — is authorization on the exported volume as strict as on ordinary reads?
• An old/duplicate path left behind after the feature's refactoring (e.g. the old endpoint was not removed, it just stopped being called from the frontend — it is still reachable and not updated to the new authorization logic).

SEVERITY SCALE (shared with the full repository audit)

A shared scale is needed so findings are comparable across individual feature checks and full audits (see security-audit-full).

• Critical: an unauthenticated external attacker achieves full compromise through this feature (RCE, access to all data of all customer companies, full authentication bypass).
• High: an authenticated user (including one with a minimal role) gains access to someone else's data/privileges through this feature — including cross-tenant access — or an unauthenticated attacker gains access to the data of one company/one user.
• Medium: requires specific conditions (a race, a particular role, MITM, social engineering) or is limited to metadata leakage/DoS without data loss.
• Low: a best-practice violation with no direct exploitation scenario at the time of the check.

For each finding, state who can exploit it (anonymous / authenticated user / internal network only) and what is lost (read / write / full compromise).

REPORT FORMAT

1. Executive summary (no technical jargon): is the feature secure to release, what is critical, what is a business/regulatory risk, what to do first.
2. SCOPE — the final list of reviewed files/directories/services (see the "Input" section above) and an explicit note on what was left OUTSIDE the SCOPE and why (e.g. "shared library X was not re-reviewed since it did not change in this feature").
3. Feature verdict: "ready to release" / "ready with caveats (see low/medium)" / "not ready — there are critical/high findings" — in one phrase at the top of the report.
4. Full list of findings: file:line, category (may reference OWASP Top 10 / OWASP API Security Top 10 / CWE), a concrete exploitation scenario, severity with justification, status, remediation recommendation.
5. A "what was done well" section — strong patterns in the feature's implementation worth replicating.
6. Action plan: what blocks the release now (critical/high), what can be fixed after release with a ticket (medium/low).
7. A "what was not checked" section — coverage limitations (no access to the prod environment, no ability to run a scanner, etc.), so the absence of findings does not read as "everything there is clean".

FINDING FORMATTING RULES

Before starting, check whether a report for this same feature already exists in docs/bugs/security_audit/ (e.g. by its feature-slug or ISSUE-ID from a previous run of this same skill). If it does, do not start numbering from zero: continue the existing ID sequence and update the status of already known findings ("not fixed" → "fixed", etc.), rather than re-raising them as new.

For each finding, the following are mandatory:

• A stable finding ID (e.g. SEC-<ISSUE-ID or feature-slug>-001), unique within the reports for this feature (see above on re-runs).
• The file path and line number (or range).
• The vulnerability name and category (OWASP Top 10 / OWASP API Security Top 10 / CWE).
• A concrete exploitation scenario: "if you make request X with parameter Y, the system will return/do Z" — not abstract wording like "there may be a vulnerability".
• Severity with justification (who can use it, what is lost).
• A remediation recommendation — concrete ("add an else branch with an explicit denial", "enable flag X by default", "replace the f-string with a parameterized query").

RUNNING THE CHECK (practical instructions)

1. First, YOURSELF (in the main thread), do the "Input" section — determine the input type (directory/branch/diff, document, YouTrack issue) and build the SCOPE. Do not delegate this step: a subagent starts without the conversation context and does not know what was meant by "the feature". Record the SCOPE explicitly before moving to the review.
2. Classify the feature by change type (new API, new UI, integration/webhook, file upload, infrastructure, etc.) and select the applicable checklist blocks.
3. Check whether a report for this feature already exists in docs/bugs/security_audit/ (see "Finding formatting rules") — this is a cheap check that saves rework and preserves the continuity of finding numbering.
4. Run the focused automated checks of PASS 1 on the files/dependencies in SCOPE (semgrep/bandit focused, SCA on the changed lock files, gitleaks on the feature's diff).
5. Perform the manual line-by-line review of PASS 2. If the SCOPE spans several services/directories and the Agent tool is available, split it into independent zones and launch a separate subagent per zone (in foreground, if the result is needed for a further decision in this same dialogue) so as not to cut corners across the whole volume at once. Give each subagent the concrete paths and the applicable sections of this skill (checklist, severity scale, finding format) — a subagent does not see this file itself. Record confirmed findings in an intermediate file as you review each zone, rather than keeping them only in context until the final report.
6. Perform PASS 3 — the feature's points of contact with the platform's existing security invariants (authentication, multi-tenancy, internal-API, shared libraries).
7. Consolidate all three passes into a single report per the format above, remove duplicates, but do not merge findings of different natures (a scanner found a pattern ≠ manual review confirmed exploitability — record both facts if both exist). Save the final report as a file at docs/bugs/security_audit/<feature-slug>-security-review.md (slug — by ISSUE-ID or by feature/directory name) — this is the same place where the registry of past audit findings lives, and the next run of this skill on the same feature should find and update it, rather than recreating it from scratch.
8. Explicitly state which checks were NOT performed (no access to prod secrets, no access to runtime logs, no ability to run a scanner, etc.) — this is part of an honest report, not its weakness.
9. If the project has a security-review skill/agent configured (a quick review of pending changes on the current branch), you can use it as a starting point for PASS 2 on the changed files, but do not limit yourself to the diff: the feature may rely on existing code that did not change in the current branch but participates in its security model (see PASS 3).

This is testing, not implementation: the developer makes the changes based on the report, not you within this skill.