Prompt injection breaks typed decision models for ~50 cents, Check Point finds
evilsocket · x · 2026-09-24
Check Point researchers show that Jev, a new class of AI model returning typed decisions instead of text, is broken in every configuration they tested: risk downgraded and investment advised on a red-flagged document, at 50 cents per successful break. Structured input didn't help; reasoning was the strongest measured defense — and Jev has no reasoning setting.
More from Models
- ChatGPT Plus quietly gets usage limits and credits, no workaround but waiting — Exarchias · 2026-09-24
- Frontier model progress is outpacing AI benchmarks, author argues — knowerofmarkets · 2026-09-24
- Xiaomi's MiMo-V2.6-Pro tops open-weight models: 46 on AA Index, 15x cheaper than GPT-5.6 — kimmonismus · 2026-09-24
- ChatGPT's Advanced Voice and Standard Voice quietly vanish from the web app — shaunralston · 2026-09-24
- DeepSeek Harness Ships Official Desktop Clients for Windows and Mac, No CLI Needed — op7418 · 2026-09-24
- DeepMind chief confirms Gemini 4 arriving in 'not so late 2026' — Novel-Nature-7741 · 2026-09-24