AI Wiki · Mediocre Security
CompanyAI security tooling

Cantina

Spearbit-derived security firm that released Apex Flash-1, a 321B-parameter open-weight model fine-tuned for vulnerability research, plus an "abliterated" variant with reduced refusals.

Spearbit-derived security firm that released Apex Flash-1 (Oct 1-2 2026), a 321B-parameter open-weight model fine-tuned for vulnerability research, plus an “abliterated” variant with reduced refusals. Scored 66.7% on 60 held-out tasks at about $2.38 per run vs $74.68 for Claude Opus 5 High (Runtime Wire, Oct 2026).

Security and assurance angle

The reduced-refusal variant is a dual-use risk.

Connections

Cantina and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

What links here

Model family Claude

Concept Fine-tuning · Jailbreak · Open-weight vs. closed models

Publication RuntimeWire

Sources

  1. Runtime Wire, “Cantina releases open-weight security model trained on 50 vulnerability cases”, Oct 2 2026