Open-weights Hy4 filters offensive security data — is open-source safety leaving defenders weaker?
douglovefishing12 · reddit · 2026-11-22
The Hy4 open-weights model shipped this week with a safety card. The team said security training focuses on defensive tasks, vulnerability analysis and remediation; exploitation-focused data was excluded, paired with refusal guardrails. By their evals, Hy4 is strong at vulnerability discovery but weak at exploitation, and they admitted over-refusal remains an issue. Pushback in the thread: chaining the model into a harness gets the same results anyway, and vuln discovery isn't inherently offensive since IR teams use it for malware analysis and forensics.
More from Models
- Screenshot of OpenAI UI seemingly showing 'GPT-6.2 Astra Ultra' goes viral — tekbog · 2026-11-22
- Yau Shing-Tung Hails Claude-Assisted Proof of Riemann Hypothesis as World-Shaking — 新智元 · 2026-11-22
- Seeking local TTS to read AI writing projects aloud on 16GB VRAM — PepSakdoek · 2026-11-22
- First Drive: Tesla FSD V15 Shines With Reality View Birds-Eye Map and Garage Auto-Parking — Scobleizer · 2026-11-22
- Grok Bot Drives 40.4% of Coinbase's Agent Trading Volume, Beating ChatGPT and Claude Combined — XFreeze · 2026-11-22
- Defending GPT-6.2: a detail-obsessed 'trees' model built for agents, not prose — koltregaskes · 2026-11-22