DeepSeek V4 and Kimi 3.1 Pro flub prompt injection help as commenter praises Claude's 'scary good' theory of mind
teortaxesTex · x · 2026-10-26
A user reports that DeepSeek V4 and Kimi 3.1 Pro are both terrible at helping craft prompt injections, with weak theory of mind, after hitting refusals from Claude Opus 5.5 and Sonnet models.
teortaxesTex attributes this to lack of training and argues Claude's theory of mind is "scary good," suggesting Dario Amodei is building for a "post-hacking era" where free will becomes the biggest vulnerability.
More from Fun
- What do you call a product you can't market but keep adding features to? — rakyll · 2026-10-26
- Grok bots can now call each other, hinting at AI-mediated socializing — xiaohu · 2026-10-26
- He told his ex the girl on FaceTime was an AI — they figured it out — Kuprel · 2026-10-26
- 'They don't know the internet is under attack': the AI security anxiety meme going around — AccBalanced · 2026-10-26
- "Monosemanticity is a pathway to many abilities some consider to be unnatural" — ZeroStateReflex · 2026-10-26
- Typo-ridden Codex prompt becomes 'made worse devisons' and still works — gowthami_s · 2026-10-26