Skip to content

fix: use configured model for API validation instead of hardcoded ret…#102

Open
audacityrifatjahanazad wants to merge 3 commits intoanthropics:mainfrom
global-competition-world:main
Open

fix: use configured model for API validation instead of hardcoded ret…#102
audacityrifatjahanazad wants to merge 3 commits intoanthropics:mainfrom
global-competition-world:main

Conversation

@audacityrifatjahanazad
Copy link
Copy Markdown

…ired claude-3-5-haiku-20241022

audacityrifatjahanazad and others added 3 commits April 20, 2026 15:03
The previous prompt told Claude to "focus ONLY on security implications
newly added by this PR" and excluded "secrets stored on disk" on the
assumption that another tool would catch them. In our setup GHAS secret
scanning is disabled, so nothing was catching them — and pre-existing
issues in touched files were silently ignored.

Changes:
- Scope: report any clear vulnerability reachable via files touched by
  the PR, not only newly-added lines. The PR is the trigger to look,
  not the limit of what to report.
- Hardcoded secrets are now in scope (API keys, JWT/RSA keys, AWS keys,
  payment-provider keys, internal API keys) regardless of file type.
- Added explicit categories that match this codebase's threat surface:
  CORS misconfig, JWT validation gaps, IDOR on wallet/payout, weak
  cert validation, dangerouslySetInnerHTML, prototype pollution,
  CloudWatch PII logging, overly permissive IAM in IaC.
- Allow LOW severity findings (was HIGH/MEDIUM only). Hardcoded test
  credentials and minor info disclosures are now reported.
- Confidence floor lowered from 0.7 to 0.6 so probable issues with
  caller-context-dependent exploitability surface for review.
- Direct Claude to use Read/Grep tools to follow data flow into
  adjacent files when needed.
- Kept the JSON output schema unchanged so downstream parsing,
  filtering, and PR-comment templates are unaffected.

All 169 unit tests still pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ew-coverage

fix(prompts): broaden scan scope so PR reviews catch real findings
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant