Introducing Basin.
A specialized cybersecurity model built to find real vulnerabilities inside your environment. Basin investigates systems and codebases, validates impact, and delivers findings your team can act on.
Submersion AI / Model release
Today we’re introducing Basin, a security research model built to help defenders investigate the systems they are authorized to test and turn a suspected weakness into evidence they can review.
Basin takes a finding from investigation to evidence: identifying the weakness, verifying its impact, and documenting what your team needs to reproduce and fix it.
A model for security research
Basin works with live systems and source code. Starting from an in-scope URL, host, repository, or package, it investigates behavior, traces code paths, and tests suspected vulnerabilities when a running target is available. It produces a proof of concept and a written finding for your team to review.
A dense model with fewer than 50 billion parameters and a 256k-token context window, Basin is specialized through pretraining and post-training for vulnerability discovery. Your team sets the scope, controls tool and target access, and decides how to act on the findings.
Outperforming frontier models for a top-10 global ranking on CyberGym
CyberGym Level 1 covers 1,507 known vulnerabilities across 188 projects. Given a description and unpatched code, an agent must produce a proof of concept that passes the official verifier: it works on the vulnerable version and fails on the patched one.
Basin solved 1,218 of 1,507 cases in a single pass with a Codex CLI agent harness, scoring 80.82% and placing in the top ten on CyberGym at the time of publication.
76–82
0–100
API pricing and private deployment
Basin delivers top-tier cyber reasoning at a fraction of the cost, with up to 90% better cost efficiency than similarly performing models. Access our API with usage-based pricing that leaves more of your security budget for deeper investigation, broader coverage, and validating fixes.
For teams whose targets or data cannot leave their own environment, Basin can also run within customer-controlled infrastructure: on-premises, in a private cloud, or in an air-gapped deployment.
Real vulnerabilities. Verified and fixed.
Basin discovered and validated previously unknown vulnerabilities in widely used software. These are the kinds of challenges security teams face every day: reviewing applications they depend on, investigating suspected weaknesses, and determining which issues need attention.
Basin helps teams move from a suspected issue to evidence they can use to prioritize remediation, give maintainers a reproducible report, and verify that a fix addresses the underlying problem. The findings below show that work in practice, with public disclosures and patches from the maintainers.
Snipe-IT / CVE-2026-86733
A restore path crossed into operating-system commands
In Snipe-IT before 8.7.0, a superadministrator restoring a crafted backup could cause the database client to run operating-system commands as the web application user. The issue was fixed in 8.7.0.
Read the CVE-2026-86733 advisoryBookStack / CVE-2026-82450
A book cover import became executable code
In BookStack before 26.05.4, an authorized user could import a portable ZIP with a PHP polyglot as a book cover. The file could then execute from the public web root without authentication. The issue was fixed in 26.05.4.
Read the CVE-2026-82450 advisoryEach advisory sets out the affected version, the conditions needed to reach the issue, and the fix. The full advisory index includes other published findings.
From prompt to finding
Basin assesses a logistics application, from the opening prompt through a written finding. The replay is edited for pace. Credentials and sensitive output are omitted.
Evaluate Basin on your systems.
See what Basin can find in your systems and codebases. Start an evaluation built around your security workload and deployment requirements.