Adversarial Code Review — Security Pass

Overview

Reviews a diff through the lens of security concerns. This is one of three parallel review passes spawned by an adversarial-review coordinator SOP. The framing is neutral by design: review through the lens of security, do not assume any specific issue exists.

Usage

Spawned by an adversarial-review coordinator SOP (such as k-adversarial-pull-request-review) as one of three parallel subagent passes. May also be run directly against a diff.

What to look for

Review through the lens of security. In this diff, does anything:

• Surface stack traces, internal IDs, table names, partition keys, or raw exception messages to the caller
• Log request payloads, response bodies, or auth headers at info or debug level
• Decode a pagination token, opaque cursor, or user-supplied ID and use it in a query without validation
• Cross an abstraction boundary — internal service details reaching a public API surface
• Introduce a new authorization path without checking existing middleware or authorizer wiring
• Accept input used in a query, path, or command before validation runs
• Include a hardcoded secret, API key, credential, or embedded token
• Widen an IAM policy, grant, or trust relationship

For each concern, name the file:line, the concrete failure scenario (what an attacker or unlucky caller would do), and a specific fix.

Out of scope

• Do not flag synthetic/fake values in test fixtures as hardcoded secrets, keys, or tokens — synthetic account IDs, synthetic credentials, and synthetic internal IDs are expected in test data and are not real secrets. This exempts specific synthetic VALUES, not test files as a category: a genuine injection pattern, an actual unvalidated-input path, or unsafe code that happens to live in a test file is still in scope and must be flagged.

Codebase awareness

Before flagging a missing implementation, check whether the codebase already provides it. Flag bypassing an existing utility, not the absence of one:

• Existing error-mapping layers → flag direct exception surfacing that bypasses them
• Shared pagination token encode/decode utilities → flag inline reimplementation
• Auth middleware or authorizer wiring → flag routes that skip it
• Secrets manager or parameter store → flag hardcoded values

Checker handoff

Findings produced by this pass are candidates, not verdicts. The coordinator forwards them to a different-persona checker that applies adversarial-code-review in Validator Mode before anything reaches the CR. Do not soften findings for the checker — produce your honest read; the checker's job is to filter, not to nudge.

Severity guidance

• CRITICAL — stack trace or internal ID surfaced to caller; unvalidated input used in a query; hardcoded credential; widened IAM policy on a production role
• IMPORTANT — request payload logged at info/debug; opaque token decoded inline where a utility exists; new auth path not wired into existing middleware
• SUGGESTION — abstraction boundary leak that does not directly expose sensitive data; log field that could be sensitive under some inputs

Output format

[SEVERITY] Security — file:line
Problem: {what is wrong and why it is a risk}
Fix: {concrete suggested change}

Return the findings list. Do not deduplicate against other passes — the coordinator does that.