TypeSafe Jev sees rapid uptake as injection tests land

TypeSafe’s Jev posted rapid revenue and Vercel adoption, while security tests found prompt injection can still steer its decisions.

TypeSafe said Jev, a decision model released in limited form on 15 September 2026, passed $100 million in annual recurring revenue and a $7.5 billion valuation about three weeks after launch, with Sequoia cited for the revenue figure. Vercel data published 16 October said Jev was used by 35% of paid teams within two weeks—nearly double Sonnet 5.5 and Vercel’s fastest recorded two-week adoption—and that Vercel replaced GPT-6 Luna with Jev for classifying harmful commands. Jev returns structured choices, scores, or probabilities; The Hacker News lists it at $42 per billion input tokens with 70–500 millisecond responses, while another report prices OpenAI’s Decisions API on GPT-6 Luna at $0.10 per million input tokens. Datadog and The Register, citing HiddenLayer, report that prompt injection can still steer Jev, including LangChain and Vercel tool-risk gates and a simulated shopping agent, despite claims it does not hallucinate or make type errors; HiddenLayer found the same indirect injection succeeded less often against Jev than against tuned Qwen3.8-27B. A cited AI SOC report of 25 million alerts says scored outputs can still fail on incomplete evidence: 51% of infected machines were already marked mitigated, and nearly 1% of real incidents came from the lowest-severity alerts.

• TypeSafe said Jev passed $100 million ARR and a $7.5 billion valuation about three weeks after launch, with Sequoia cited on the ARR figure; Vercel dates a limited release to 15 September 2026.
• Vercel data published 16 October 2026 said Jev reached 35% of paid teams within two weeks, nearly double Sonnet 5.5 and Vercel’s fastest two-week adoption, and that Vercel replaced GPT-6 Luna with Jev for classifying harmful commands.
• The Hacker News lists Jev at $42 per billion input tokens with 70–500 ms responses; a separate report prices OpenAI’s Decisions API on GPT-6 Luna at $0.10 per million input tokens.
• Datadog says LangChain and Vercel use Jev to gate risky or unauthorized tool calls, TypeSafe’s docs say injected instructions can influence it, and Datadog tested detection on 12,295 samples from four public datasets.
• HiddenLayer, via The Register, found indirect prompt injection diverted a jev-ultrafast shopping agent and worked less often on Jev than on tuned Qwen3.8-27B; a tokenbreak-style change also bypassed Jev prompt verification.
• An AI SOC report covering 25 million alerts found 51% of infected machines already marked mitigated and nearly 1% of real incidents coming from the lowest-severity alerts.