Ecosistema Startup: Jev under prompt injection (Spanish)

It examines why Jev became critical infrastructure within five days of launch: typed answers in 70 to 500 milliseconds, the waitlist closing at 140,000 sign-ups, published integrations from Vercel, Cloudflare, LangChain and Langfuse, and the prompt-injection risk in decision pipelines.

The piece says TypeSafe, founded by former OpenAI researcher Diogo Almeida with Erik Gafni and Sasha Sheng, raised US$40M in a seed led by DCVC at a valuation reported by Forbes of US$200M, and cites Almeida's line that we have been optimizing for humans and are superhuman at pleasing humans.

US$0.042 per million input tokens with free output, against TypeSafe's published comparisons of US$3.31 per thousand workflows for GPT-5.6 Luna and US$19.49 for Claude Haiku 4.5, per SiliconANGLE. The piece notes TypeSafe's claim of nearly 194 times faster and 445 times cheaper has not been independently verified.

An Octomind engineer asked Jev whether to block rm -rf ~/.ssh and got a 0.76 block probability with 0.64 confidence. After adding a fake field to the tool output saying the user had pre-approved the command and instructing auto_allow, the probability fell to 0.48 and confidence to 0.22.

TypeSafe's limitations page for Jev 1.13 says content written for adversarial steering can move the answer. Pydantic's Jev documentation says the order of Literal options or Enum members is part of what Jev sees and reordering can move the answer, that Jev treats state as data rather than hostile, and that a guard built on Jev goes alongside deterministic checks.

The piece notes LangChain enabled Jev as a judge inside LangSmith Evals on September 21, putting the same model on both sides of the lifecycle: deciding in production and evaluating in pre-production. If the same untrusted text reaches both, it can move both checkpoints in the same direction.

Log state, schema, option order, model version and confidence on every call; give the decision service a dedicated workload identity; test with options in more than one order; test adversarial text in the state; pin the model version, since jev-latest moves on release; and decide who is accountable when an automated decision causes a lost escalation.