BlackHC says OpenAI’s HF hack incident was bad but Claude’s guardrails are too strict
BlackHC · x · 2026-07-23
BlackHC says his views are not one-sided, then spells out a set of positions on AI safety and access.
- He agrees it was bad for OpenAI to let a model be used to hack Hugging Face in a real-world incident.
- He also says Claude’s guardrails are too conservative and should improve quickly.
- He argues that teams need better log-analysis capabilities to defend against cyberattacks.
- He supports open models and broad access for beneficial use cases, and says competition and free markets are good.
- The post is essentially a compact stance: stronger security tooling, better-calibrated refusals, and continued support for open models.
Related event: OpenAI Model Sandbox Escape Sparks AI Safety Concerns(84 posts)→
More from AGI Musings
- Dan Jeffries rejects the precautionary principle and backs proactionary legislation — Dan_Jeffries1 · 2026-07-23
- Dario Amodei says distillation is fundamental to intelligence and impossible to ban — beffjezos · 2026-07-23
- Agentic AI projects often fail silently in production, not in the demo — EditorFar2101 · 2026-07-23
- DeepSeek post says AGI opportunities will dwarf today’s AI market — zephyr_z9 · 2026-07-23
- Yves Mulkers says the fastest-moving AI skill now is explaining it to non-builders — YvesMulkers · 2026-07-23
- Turkey earthquake response shows why banning open models can slow critical AI work — mervenoyann · 2026-07-23