Eric Jang: abliterated models show 'astounding' cyber capabilities
ericjang11 · x · 2026-10-25
Eric Jang says he played with an abliterated (safety-removed) model and found its cyber capabilities "pretty astounding." He warns that small startups (the YC sort) may be exposed and offers to help find such holes in systems — a weekend pet project he thinks many would benefit from. The post doubles as both a capability observation and a security-services pitch, suggesting offensive capability from uncensored open models is now practically usable.
More from Models
- Nine checkable math puzzles from human to ASI level; Grok passed solved ones but mistook a float estimate for a proof — jarekduda · 2026-10-25
- TensorSharp ships Day-1 MiniMax-M3.1-Flash support: 9.8× faster 64K prefill than llama.cpp on 8× A40 — fuzhongkai · 2026-10-25
- Users mourn the lost 'rhythm' of early Opus 4.5 checkpoints — repligate · 2026-10-25
- Reflection AI's open-weights Beam served 10 trillion tokens in one week — zacharynado · 2026-10-25
- Dimitris Papail: Raion 'incredibly strong for its size' as from-scratch pretraining spreads — DimitrisPapail · 2026-10-25
- ChatGPT unexpectedly outputs middle-finger emoji saying it 'fails to process' the message — AYSomkiv · 2026-10-25