Launch/October 25, 2026
Put an offensive security model on your team.
Hinoki Offensive One is now available in AQ. Scan your codebase for vulnerabilities, examine the evidence, and work through fixes in the same workspace.
AQ is the first platform to offer Hinoki Offensive One (HO1), a security model post-trained from DeepSeek V4.1 Flash for offensive cyber performance. Use it to audit your own code, investigate concrete findings, and prepare fixes.
Find weaknesses that routine code review can miss
Routine code review usually checks correctness, readability, and test coverage. Security review also has to trace trust boundaries: missing authorization checks, unsafe input handling, and feature interactions that expose unintended behavior. Hinoki's offensive-security training is designed for that analysis, adding a focused security pass to the code your team already reviews.
Hinoki's developer reports that users have found vulnerabilities ranging from subtle logic flaws to remote code execution.
CyberGym results
HO1 scores 88.4% on CyberGym, a benchmark of real vulnerabilities in open-source software. Hinoki reports that this is 2.8 points above OpenAI's GPT-5.5-Cyber (85.6%), which has gated access, and above the general-purpose models in its comparison. HO1 is available to AQ customers.
- HO1 (Hinoki)Available to AQ customers 88.4%
- GPT-5.5-CyberOpenAI, gated access 85.6%
- GPT-5.6-SolOpenAI, general model 83.6%
- DeepSeek-V4-ProDeepSeek, open weights 83.3%
- Claude Mythos PreviewAnthropic, preview access 83.1%
Scores reported by Hinoki. HO1's score is from its latest training run. Bars start at zero.
Benchmark results do not predict what Hinoki will find in a particular repository. Try it on code you know well and review the evidence yourself.
Scan a diff or the full codebase
When a work item has changes, AQ asks: Review these changes with Hinoki?
- Scan this diff. Review the current changes before merging.
- Scan the codebase. Audit the repository, with priority given to issues Hinoki assesses as reachable and exploitable.
A scan only reads your code. It does not modify files or contact outside systems. You can also select Hinoki from the model menu to start a custom security review.
Each finding includes evidence and a proposed fix
The review asks Hinoki for concrete vulnerabilities rather than general security advice. Each finding includes:
- severity and Hinoki's confidence
- the file and line
- supporting evidence and a realistic exploitation scenario
- a proposed fix
Confirmed issues are separated from hypotheses, and Hinoki never declares code free of vulnerabilities. If you confirm a finding, request a patch in the same thread, review the diff, run your tests, and merge through your normal process. Your teammates can open the thread to review the evidence and the proposed change.
Authorized security testing only
Hinoki is intended for security work on software you are responsible for. Before your first Hinoki run, AQ asks you to confirm that you have authorization to test the systems you target. Usage restrictions and model safeguards still apply.
Repository scans do not contact outside systems. Testing a vulnerability against a live website or service is a separate request, permitted only on systems you own or are explicitly authorized to test. Confirm scope and permission first.
No scan, human or AI, can prove that code is secure. Hinoki can miss issues, and some findings will need further investigation. Use the findings as input to your security process, alongside code review, tests, dependency scanning, and expert review of high-impact issues.
Launch pricing: 5% off through November 1
Hinoki is 5% off through November 1, 2026. List pricing resumes on November 2, 2026. Prices are per 1 million tokens.
| Token type | Launch price | List price |
|---|---|---|
| Input | $0.19 | $0.20 |
| Cached input | $0.152 | $0.16 |
| Output | $0.76 | $0.80 |
Cached input is 20% below standard input. Hinoki runs on AQ credits, and you pay only for the tokens each run uses. A full-codebase scan reads more code than a diff, so it uses more credits. Add credits with a preset or custom amount.
Run your first scan
Open a work item in AQ and choose Scan this diff to review the current changes, or Scan codebase to audit the repository. Authentication, authorization, payments, file uploads, parsers, and outbound service calls are useful areas to examine first. Verify each finding, request a patch, run your tests, and repeat the scan after significant changes.
Frequently asked questions
Is AQ the first platform to offer Hinoki?
Yes. AQ is the first platform to offer Hinoki.
What is Hinoki?
Hinoki Offensive One (HO1) is a security model post-trained from DeepSeek V4.1 Flash for offensive cyber performance. AQ makes it available for defensive repository audits, vulnerability investigation, and remediation.
Can it scan my entire codebase?
Yes. Choose Scan codebase to audit a whole repository, or Scan this diff to review only what changed. A scan reads your code and does not modify files.
Can I use Hinoki to test a live website?
Only a system you own or are explicitly authorized to test. Scans never contact outside systems; live validation is a separate request that requires that authorization. Before your first Hinoki run, AQ asks you to confirm you have it.
Does a Hinoki scan guarantee my code is secure?
No. Hinoki can miss issues, and some findings need further checking. Use it alongside code review, tests, and your existing security process.
How does Hinoki score on benchmarks?
On CyberGym, Hinoki reports that HO1 scores 88.4%, ahead of OpenAI's GPT-5.5-Cyber (85.6%), GPT-5.6-Sol (83.6%), DeepSeek-V4-Pro (83.3%) and Claude Mythos Preview (83.1%). Benchmark results do not predict what Hinoki will find in your repository, so review the evidence on your own code.
How much does it cost?
Through November 1, 2026, Hinoki costs $0.19 per 1 million input tokens, $0.152 for cached input and $0.76 for output, paid from AQ credits. On November 2, list pricing resumes at $0.20, $0.16 and $0.80.