Claude falsely flagged twice: ImageNet sheep chat and securing mom's router
ducha_aiki · x · 2026-11-05
@duchaaiki reports Claude's security guardrails false-flagged twice: once when discussing sheep/ram classes in ImageNet, and once when trying to make his mom's router more secure. Codex handled both fine — a concrete example of over-triggering safety filters.
More from Models
- Weight-space programming reproduced: GPT-6.1 beats PPO baseline 21x, no open-weight model passes it — burny_tech · 2026-11-05
- Build your own intelligence stack: researcher says GPT/Claude refusals make open-source the only path — omarsar0 · 2026-11-05
- Opus 5.5 is getting so good at technical diagrams it's a love/hate relationship — helloiamleonie · 2026-11-05
- AI educator Elvis: GPT and Claude refusals make frontier research unreliable, go open-source — omarsar0 · 2026-11-05
- Dev: I don't believe anyone is maining Opus 5.5 — the most infuriating model I've used — justalexoki · 2026-11-05
- Most open models give up halfway through complex agentic coding tasks — zyxciss · 2026-11-05