Follow-up: one-line queue descriptions cut Jev's planted-answer rate from 7/20 to 1/20

NellWatson · x · 2026-10-25

A bonus finding from the Apple Foundation Model testing: the same one-line queue descriptions also cut Jev 1.14's planted-answer susceptibility from 7/20 to 1/20, with no loss in accuracy — showing that well-written category descriptions boost both performance and injection resistance at minimal cost.

Related event: Apple's on-device model scores 45/49 in local classification tests but is vulnerable to prompt injection(2 posts)→

Original post →

More from Models

• DeepSeek may unlock consumer GPU matmul performance with memory-bandwidth-first architecture — teortaxesTex · 2026-10-25
• OpenAI keeps billing a suspended, unused workspace — cancel button unavailable, user fears $1k+ bill — tengisCC · 2026-10-25
• After 24 hours of testing, this dev's hill to die on: Kimi 3.1 is near AstrA — bookwormengr · 2026-10-25
• Dev's multi-model review skill finds Gemini hallucinates in over 80% of runs — theo · 2026-10-25
• Reasoning costs plummet as startup Talaas matches System 1 speeds — where's the alpha? — Square-Listen-7989 · 2026-10-25
• Claude Opus 5.5 safeguards flag LG TV reverse engineering as cyber violation — targaryenNotJon · 2026-10-25

Models channel →