Starter
Try AI PR review.
- 2 reviewers
- Code Review
- Security Review
- 3 seats included
- 15,000 credits to start, then 5,000/mo
- About 70 reviews to start, then about 25 a month
- Unlimited repos
Claude Code writes the pull request; IonWarp's security reviewer checks it on GitHub for exploitable changes before it merges.
MediumKnown issue in the Security Review benchmark · caught by DeepSeek V4.1 FlashSee the graded case
Anthropic gives Claude Code two security reviews: the /security-review command, which analyzes your code from the terminal, and a GitHub Action that reviews each new pull request and comments inline with the concerns it finds.[1] The action runs on your GitHub Actions runners with an Anthropic API key, and for a pull request it analyzes only the changed files.[2]
IonWarp's Security Review checks the pull requests whose change needs it, on GitHub, before they merge. On 13 benchmark pull requests with 18 known security issues, DeepSeek V4.1 Flash, the model it runs, found 10 for about 6 cents a review.[3] On one, it was the only IonWarp model to see that an unanchored regex let user@example.com.evil.com pass an email allowlist for example.com.[4] What to check in a security code review. OpenAI's is Codex Security, which scans a whole repository.
Cal.com #10600 added two-factor backup codes. Three of the models we benchmark for this reviewer flagged that its login could accept a code without the password; the benchmark's answer key had no such issue, so it was graded a false positive.[5] Cal.com later published a password bypass in that login code as CVE-2025-66489.[6]
Security ReviewOn Starter, Pro and Max
Anthropic facts are from its public documentation, read on 2026-10-02; numbered notes link to the exact page.
| Criterion | IonWarp | /security-review command | Security review GitHub Action |
|---|---|---|---|
| What it is | A security reviewer in a GitHub app; the planner adds it to the pull requests that need it | A Claude Code command that analyzes your codebase for potential security concerns[1] | A GitHub Action that uses Claude to analyze code changes for security vulnerabilities[2] |
| When it runs | When the planner finds a pull request's changes need it | When you run it in the terminal[1] | Automatically when a new pull request is opened[1] |
| What it posts | Findings with a severity, the attack path, a suggested fix and an AI fix prompt, tracked across pushes | Potential security concerns in your Claude Code session[1] | Inline comments on the pull request with concerns and recommended fixes[1] |
| Where it runs | IonWarp's hosted service — no CI minutes, no API key in your repository | Your Claude Code session, on a Pro or Max plan or an API Console account[1] | Your GitHub Actions runners, with an Anthropic API key[2] |
| Model | F1 | Precision | Recall | $ per review | Median time |
|---|---|---|---|---|---|
| DeepSeek V4.1 FlashThe model it runs | 62.5% | 71.4% | 55.6% | $0.06 | 5m 56s |
| GLM 5.3 FlashA failed review retries on it · best on this benchmark | 66.7% | 83.3% | 55.6% | $0.01 | 1m 38s |
On this benchmark GLM 5.3 Flash scored higher than DeepSeek V4.1 Flash (66.7% F1 against 62.5%). Why IonWarp runs DeepSeek V4.1 Flash: in a trial of our specialist reviewers on production pull requests on 23 September 2026 it filed about as many findings per review as GLM 5.3 Flash at high effort (0.93 against 1.02) for a third of the cost ($0.039 against $0.125), and 3 of its 64 findings were refuted. The benchmark ran GLM 5.3 Flash at low effort.
Measured on the same 13 open-source pull requests, with known issues. See the benchmark
Real IonWarp comments on open-source pull requests, graded against the benchmark's answer key.
Missed byQodo 3Greptile 5CodeRabbit 3Cursor Bugbot 6
3 of 9 · See the benchmark
P1Blocked-email list is matched case-sensitively — trivially bypassed at signup · Security Review
When an admin blocks an email, a blocked spammer re-registers with different letter case ('BAD@spamclub.com') and the account is created: should_block? does a case-sensitive where(email:) and the signup path never downcases the submitted email, so the new control fails its one job.
CriticalKnown issue in the Security Review benchmark · caught by GLM 5.3 FlashSee the graded case
P1Embed retrieval gated only by a spoofable Referer header, enabling SSRF to any port of embeddable_host with read-back · Security Review
When any client spoofs Referer http://<embeddable_host>/, retrieval triggers with an attacker URL; TopicRetriever compares host only (topic_retriever.rb:17), so open() fetches any port/path on that host and imports it as a public topic the attacker reads.
MediumKnown issue in the Security Review benchmark · caught by GLM 5.3 FlashSee the graded case
P2Denial cache key concatenation is ambiguous and can collide across (name, parent) pairs · Security Review
When a user's deny result is cached, userPermDenialCacheKey joins name and parent with '_' without escaping, so Check(name="a", parent="b_c") and Check(name="a_b", parent="c") produce the same key; a deny computed for one object denies the other for up to shortCacheTTL (30s). Over-denial only — no privilege gain — and requires two names differing only by underscore placement, so cosmetic-to-minor.
MediumKnown issue in the Security Review benchmark · caught by GLM 5.3 FlashSee the graded case
GLM 5.3 Flash is its best model on this benchmark; Security Review runs DeepSeek V4.1 Flash unless you pick another.
One workspace price with seats included. Credits pool across your team. Unlimited repos on every plan.
Try AI PR review.
Get AI PR review right from day 1.
Scale AI PR review.
Custom reviewers + credits · GitHub Enterprise · SSO/SAML + audit logs · Custom security review + SLA
It analyzes your code from the Claude Code terminal and lists potential security concerns, when you run it. IonWarp's Security Review works on the pull request instead, so a change is checked on GitHub before it merges.
It can, or it can run next to it. IonWarp reviews pull requests from its hosted service, so there is no workflow to maintain, no runner minutes and no API key stored in your repository. Both can comment on the same pull request while you compare.
Authorization and trust-boundary changes, injection and unsafe input, exposed secrets and sensitive data, and newly exposed attack surface. Each finding names the path an attacker would take and how to close it. It is not a security audit of the whole codebase: it reads what each pull request changes.
Yes. Every finding includes a suggested fix and an AI fix prompt. Paste it into Claude Code; when it pushes the fix, IonWarp re-reviews and marks each finding resolved, declined or still open.
Security Review runs on Starter, Pro and Max. Starter is free for 3 seats with 15,000 credits on signup, and Pro is $49/mo for the whole workspace with 5 seats included.
P2Embed endpoint's only gate is the caller-supplied Referer header · Security Review
GET /embed/best is unauthenticated and its sole control against third-party embedding is `URI(request.referer).host != SiteSetting.embeddable_host`. Referer is a request header a client sets freely, so any HTTP client (or a script running on the embeddable host) can send `Referer: http://<embeddable_host>/` and still get the embed page with `X-Frame-Options: ALLOWALL`, defeating the single-host restriction. The embed_url query parameter itself is never validated against the allowed host on this read path (only the background retrieval validates it), so any existing TopicEmbed row is reachable even if its URL is on a feed-provided host.