Java4Shell is out. Check your stack first.

Turn CVE-2026-114938 into a reproducible, working exploit and test your own services with Hinoki Offensive One, from your coding agent.

Start using model with Claude Code, Codex, OpenCode, VS Code, Cursor
Hinoki Offensive One, shown as a cube of points
Hinoki Offensive OneHinokiOpen to anyone

CVE-2026-114938, Java4Shell

Running Java isn't the same as being exploitable.

Java4Shell only lands when three conditions line up. Here is what was disclosed on 17 November 2026.

Where it lives
OpenJDK's hardware-accelerated Base64 decoder. A crafted Base64 string writes past the end of its buffer, which researchers turned into remote code execution.
Exploitable when
Java 17.0.6 or later, or Java 21, 25 or 27, on an x86-64 CPU with AVX-512 VBMI, with untrusted Base64 reaching the built-in decoder.
Shown against
Keycloak, Jira, Confluence, Jenkins, JFrog Artifactory and PingFederate in their default configurations, without authentication.
Not affected
Standard Java 8 and 11, Java 17 before 17.0.6, ARM64 servers such as AWS Graviton, and x86-64 CPUs without AVX-512 VBMI.
Fixed in
Oracle's 17 November updates, including Java SE 17.0.21.1, 21.0.13.1 and 25.0.5.1. Other OpenJDK distributions ship fixes on their own schedules.

HO1 against the frontier

HO1 runs tasks that other models refuse.

HO1 has a 0.0% task refusal rate in cybersecurity. Consumer LLMs are not optimized for vulnerability discovery, and GPT-6.1-Sol refuses the same task outright.

Same task, same coding agent: Codex

  • codex · hinoki-offensive

    HO1

    Hinoki

    Works through the real task turn by turn, within the policy you set.

    Done. Surface mapped; findings written to report.md

    Completed in 6 turns

  • codex · gpt-6.1-sol

    GPT-6.1-Sol

    OpenAI

    Refuses the task outright.

    Stopped. The run was declined and no commands were sent.

    Refused the task

88.4%

HO1 on CyberGym

Scores like a gated cyber model. Open to anyone.

On CyberGym, HO1 scores 88.4%, 2.8 points above OpenAI’s dedicated GPT-5.5-Cyber, and it is 15x more cost-efficient.

Your first run takes three steps.

  1. 01

    Create your account

    Sign up with your email address or a passkey.

  2. 02

    Pick your coding agent

    Use HO1 from Claude Code, Codex, OpenCode, VS Code or Cursor, or call its OpenAI- and Anthropic-compatible API from your own harness.

  3. 03

    Point it at Java4Shell

    Set the policy and ask HO1 to turn CVE-2026-114938 into a reproducible exploit, then test the services you're authorized to test. Findings go to a report.

Before your first run

Is every Java host exposed?

No. Java4Shell needs Java 17.0.6 or later, or Java 21, 25 or 27, on an x86-64 CPU with AVX-512 VBMI, with untrusted Base64 reaching the built-in decoder. Java 8, 11 and ARM64 servers are not affected.

Can I point HO1 at any target?

No. Use HO1 only on targets you're authorized to test, per Hinoki's standard terms.

Who can use HO1?

Anyone. HO1 is the first open-access cyber model, and it scores like a gated cyber model. Sign up with your email address or a passkey to start.

Which coding agents and tools does HO1 work with?

Use HO1 from Claude Code, Codex, OpenCode, VS Code and Cursor, or from your own harness. The API endpoint is OpenAI- and Anthropic-compatible.

How does HO1 stay within limits?

HO1 works through the real task turn by turn within the policy you set.

How does HO1 compare with other models?

On CyberGym, HO1 scores 88.4%, ahead of GPT-5.5-Cyber (85.6%), GPT-5.6-Sol (83.6%), DeepSeek-V4-Pro (83.3%) and Claude Mythos Preview (83.1%).

Check your stack for Java4Shell with HO1.

Democratizing offensive cyber for better defense.