Check Point Breaks Decision Model Jev for About 50 Cents per Attack
evilsocket · x · 2026-10-08
Check Point's AI security team red-teamed Jev, a new kind of model that returns typed decisions instead of text for software to consume. Every configuration they tested broke: risk was downgraded to low with an investment advised on a document full of warning signs. Each successful break cost about 50 cents, with the strongest attacker landing on the fourth turn on average. Typed, structured inputs didn't make the model meaningfully harder to manipulate, and neither did instructing it to distrust the document. Reasoning was the strongest defense measured — but Jev has no reasoning setting. The core risk: nobody reviews a decision the way they review text, making these models arguably more dangerous to attack.
More from Models
- Nous Research Launches Hermes Index to Rank Models Inside Hermes Agent — NVIDIAAI · 2026-10-08
- ByteDance Seed Paper Explains Phase Blind Spots in KV Compression Behind DeepSeek's Erratic Long- Context Performance — teortaxesTex · 2026-10-08
- Gary Marcus Slams OpenAI's Vague Math Proof Report: Zero Details, Won't Pass Peer Review — GaryMarcus · 2026-10-08
- Perplexity releases pplx-embed-v2-late: OCR-free late-interaction embeddings topping retrieval benchmarks — perplexity_ai · 2026-10-08
- Anonymous stealth LLM 'Space Bunny Alpha' tops OpenRouter with 22% usage share — maferase · 2026-10-08
- China's DeepSeek and peers shipped 16 AI models in a month — pstAsiatech · 2026-10-08