BynarIO AI

What it is

Bynario describes BynarIO as an autonomous security system for software understanding, vulnerability discovery, validation, and repair. Its public materials emphasize compiled-binary analysis, LLM-assisted vulnerability research, proof-of-concept validation, and patch generation.

For Bugflation, the important scope is narrower: Bynario now has public vulnerability records where BynarIO AI or Bynario AI is named in accepted disclosure workflows, including an Apple product-security advisory and Linux kernel commits.

What is verified

The current ledger indexes three conservative Bynario entries:

• CVE-2026-31532, a Linux kernel CAN raw socket use-after-free. Bynario’s write-up says its LLM-driven pipeline discovered, validated, and patched the vulnerability, with Opus 4.6 as the primary model for this case. The upstream Linux commit independently names Bynario AI in an Assisted-by trailer, and NVD records the kernel.org CVE as CVSS 7.8 high.
• CVE-2025-43377, an Apple Model I/O out-of-bounds read. Apple directly credits BynarIO AI (bynar.io) in the macOS Sequoia 15.7.2 security content, while Bynario’s launch material describes autonomous binary analysis of Apple’s USD library and trigger generation.
• CVE-2026-31694, a Linux FUSE page-cache overflow. The upstream Linux commit includes Assisted-by: Bynario AI, and NVD records the kernel.org CVE as CVSS 7.8 high. The same commit also carries separate Reported-by credits, so Bugflation scopes this as a direct AI-assisted accepted fix rather than a clean solo discovery attribution.

These are direct-attribution entries because the AI credit appears in upstream or vendor records, not only in Bynario’s own posts.

Attribution boundary

CVE-2026-31694 is included because Bugflation’s methodology allows accepted AI-assisted fixes, not only solo AI discovery. The public source trail names Bynario AI in the upstream commit and confirms the CVE. The caveat is narrower: the entry should not erase the separate human reporter credits, and it should not be presented as an uncontested Bynario-only discovery unless later public material clarifies that path.

There is also a source caveat on the CAN post: its remediation section links to the FUSE commit instead of the CAN commit. The correct CAN upstream commit is a535a9217ca3f2fccedaafb2fddb4c48f27d36dc. The mismatch looks like an editorial link error, not a reason to reject the finding, because the CAN commit and NVD record independently corroborate the issue and Bynario AI assistance.

Why it matters

Bynario adds another high-quality bugflation signal: direct AI attribution in accepted vendor and upstream records, plus a technical write-up describing how LLM-driven validation handled a subtle kernel race that ordinary KASAN instrumentation would not automatically catch.

The Apple entry shows Bynario’s closed-source binary-analysis story. The Linux CAN entry shows source-level kernel reasoning, validation, and patching. Taken together, they make Bynario publishable in Bugflation while still requiring careful boundaries around unindexed or mixed-credit work.

Sources

• Bynario: Discovery & Validation in the Linux Kernel, Part 1
• Bynario: An Exciting New Chapter in Security Research
• Bynario: The idea behind Bynario
• Linux commit a535a92: can raw ro->uniq use-after-free
• NVD: CVE-2026-31532
• Linux commit 51a8de6: FUSE oversized dirents
• NVD: CVE-2026-31694
• Apple: macOS Sequoia 15.7.2 security content
• OpenCVE: CVE-2025-43377