security-review ยท diff
git:20260717.e1d5341 to git:20260718.42bcc8e
10 added, 22 removed. Audit A to A.
---
name: security-review
- description: Perform a focused security review of changed code and configuration. Use when work touches authentication, authorization, secrets, command execution, file paths, archives, network calls, redirects, user input, databases, dependencies, plugins, MCP tools, logs, permissions, sandboxes, or agent boundaries.
+ description: Review a focused trust-boundary or exploit-path change without modifying it. Use when work introduces or materially changes untrusted input, privilege decisions, secret flow, dangerous sinks, or agent and tool capabilities.
---
# Security Review
- Review the security-sensitive surface changed by the current task.
+ Own concrete trust-boundary and exploit-path analysis, not generic changed-code review.
## Workflow
- 1. Identify touched trust boundaries, untrusted inputs, and dangerous sinks.
- 2. Trace external input through parsing, validation, authorization, and side effects.
- 3. Check for committed or exposed credentials and sensitive log data.
- 4. Check command construction, quoting, argument boundaries, and injection paths.
- 5. Check file paths, archives, uploads, and downloads for traversal, unintended overwrite, unsafe deletion, and resource exhaustion.
- 6. Check network behavior for SSRF, open redirects, insecure transport, credential leakage, and unbounded requests.
- 7. Check authentication and authorization for missing enforcement, privilege escalation, tenant leakage, insecure defaults, unsafe session or token lifecycle, and unauthenticated persistent connections such as WebSockets.
- 8. When cryptographic material or protected backups are touched, check key ownership, rotation, recovery copies, and exposure during backup or disaster recovery.
- 9. Check dependency and tool changes for unexpected sources, unsafe permissions, and inappropriate version pinning.
- 10. Check containers, plugins, MCP, sandboxes, and agent configuration for hardcoded secrets, broad mounts or egress, metadata-service access, token exfiltration, escape paths, and hidden side effects.
- 11. Fix confirmed issues within scope and report unrelated risks separately.
-
- ## Finding Standard
+ 1. Confirm the diff or audit scope, assets, actors, trust boundaries, and relevant privileges.
+ 2. Trace each material untrusted source through parsing, validation, authorization, and dangerous sinks or persistent side effects.
+ 3. Check changed secret handling, command execution, paths and archives, network destinations, authentication, authorization, tenant isolation, logs, plugins, MCP tools, sandboxes, and agent capabilities only where they participate in a concrete path.
+ 4. Separate confirmed exploit paths from defense-in-depth observations and unresolved evidence gaps.
+ 5. Delegate dependency provenance to `dependency-review`, IaC blast radius to `infra-review`, platform lifecycle to its specialist, and operational recovery to `rollout-readiness-review`.
- For each confirmed finding, identify:
+ For each finding report severity, affected location and behavior, attacker prerequisites, exploit path, impact, smallest safe correction, verification, and uncertainty. State directly when no confirmed finding exists.
- - affected location and behavior;
- - exploit or failure path;
- - severity and impact;
- - smallest safe fix;
- - verification performed.
+ Keep the review read-only. Do not rotate credentials, modify live services, run unsafe payloads, add security frameworks, or broaden scope. Remediation requires separate implementation authority.
- Do not report speculative concerns as confirmed vulnerabilities, rotate credentials, modify live services, add security frameworks, or broaden the task into a general security rewrite without authorization.
+ Finish when material trust paths in scope are traced and findings or confidence limits are explicit.