Replayable Code Security Review
Before a code change is merged, this tool reproduces exploitable vulnerabilities in an isolated environment and returns attack evidence, a minimal patch, and a regression test.
When a small team is about to merge AI-assisted login, upload, or authorization code, it gives the product the repository and the branch to be merged. Rather than returning a context-free list of risks, it uses the changes to generate targeted attack attempts, such as unauthorized access, path traversal, token replay, or input bypass.
Each attempt runs against an isolated copy. If an attack succeeds, the page records the request parameters, execution path, scope of affected data, and a command that can be rerun. Developers can follow the call stack to see which route, validation condition, or permission check let the vulnerability through. Alerts that cannot be reproduced reliably are moved to a review-needed area instead of occupying the first screen of the merge check.
For reproduced issues, the product proposes a minimal fix and includes a regression test that must fail on the old code and pass on the new code. Developers can edit the patch in the interface, rerun the attack case, then commit the confirmed fix and test back to the pull request.
The first release focuses on common authentication, authorization, and input-handling issues in web services, and runs only with test data in isolated environments. It does not scan production accounts or publish attack steps to public channels; review results are visible by default only to members authorized for the repository.
Why now
When observed on July 29, Codex Security ranked third on Hacker News with 304 points and 75 comments. S1 The official tool already links attack-path analysis, isolated reproduction, and minimal patches into a closed loop, so small teams will bring the same standard into merge review sooner. S2
Target user
The target user is a two-to-ten-person development team without a dedicated security engineer. They have just used AI to modify login, upload, or authorization code and are about to merge a pull request. At this point, the scope of change is still clear and the test environment is easy to rebuild. They need evidence of exploitability, not more risk labels. A fix is most likely to fit the existing review workflow when it can be committed together with a regression test.
Minimal entry point
The first release is limited to JavaScript and TypeScript web services on GitHub. A GitHub App receives pull-request events and checks out the target and merge branches. It uses the Codex Security TypeScript SDK to scan the diff, then calls its validation and patch capabilities. Each job runs in an unprivileged Docker container with only a repository copy and temporary test data mounted. For authentication issues, projects must provide test-account fixtures. Reproduced HTTP requests can be saved as Supertest tests. The interface first shows the request, call path, and rerun command. Production probing, cross-repository flows, and custom infrastructure are not supported initially.
Punching above its weight
Find initial users among small GitHub teams that use AI coding tools frequently. Publish a public demo repository with a real vulnerable branch, where an access-control issue produces opposite results before and after the fix. Then ship the checker as a GitHub App that maintainers can install on a single repository for a trial. Publishing reproduction cases and a test generator can draw security engineers into reviewing the rules. Acquisition content should show one complete merge check, not speak broadly about AI security.
Competitors & gaps
- OpenAI Codex SecurityGoogle
- Codex Security already covers most of this product’s core workflow. It builds a threat model for the repository, explores attack paths, and validates findings in an isolated environment. After validation, it can propose a minimal patch, which a person can then turn into a pull request. Its official CLI already supports diff scanning, result export, validation, and patch commands, while the TypeScript SDK provides diff targets, cancellation, and structured results. Simply wrapping a scan would therefore be hard to differentiate. The remaining opening is a narrower pre-merge experience: turn reproduced requests directly into project tests, clearly show the old branch failing and the fixed branch passing, and let developers edit patches and rerun cases in place. If those interactions merely relay Codex Security, teams will likely use the official tool directly.
- GitHub Code Security (CodeQL and AI Security Detection)Google
- GitHub CodeQL can already run static analysis on pull requests. It supports several major languages and surfaces results as checks or code annotations. Some repositories can also use AI security detection and Copilot Autofix, without leaving the existing code-review workflow. Its strengths are mature rules and permissions already integrated with GitHub. AI detection can also cover languages and frameworks that CodeQL does not. Yet these results are still primarily alerts and repair suggestions, and GitHub’s documentation notes that AI alerts can produce false positives. The opportunity is to execute a real attack and retain complete reproduction material. Each finding should also include a rerunnable command and a regression test. Without reliably producing this kind of dynamic evidence, the product becomes just another pull-request alerting system.
How it makes money
Charge a monthly subscription per private repository. Plans include a fixed scanning allowance and member seats, with per-pull-request charges for scans beyond the allowance. Offer a limited free tier for public repositories.
The case against
Platform vendors already provide similar capabilities, leaving little room to differentiate. The costly part is rebuilding an executable environment for each project: without dependencies, databases, test accounts, and secrets, an attack may not reproduce. Weak isolation can also expose source code or credentials, creating a serious trust cost. A generated patch may make one case pass without correcting the authorization model. Long scans can slow merges and encourage teams to bypass the check. This is worth pursuing only if it can first produce stable, rerunnable evidence in one technology stack.