What is TypeSafe JEV? A complete technical breakdown of TypeSafe AI’s JEV System One decision model, sensitive data leakage risks, and how Moftgate AI DLP monitors and blocks unauthorized traffic.
Quick Summary (TL;DR)
JEV is a non-generative "System One" AI model released by TypeSafe AI that makes single-pass, sub-second typed decisions instead of writing text. Because JEV ingests raw state data (customer PII, financial logs, source code) to produce classifications, it presents serious data exfiltration risks. Moftgate provides endpoint-level visibility, policy-based domain blocking, and real-time AI DLP inspection to stop sensitive leaks before they reach JEV servers.
In September 2026, TypeSafe AI announced a new model called JEV. JEV is fundamentally different from traditional generative chat models. It does not write text or generate code. Instead, it ingests unstructured program state and returns a typed, probabilistic decision in a single parallel pass. This introduces a major challenge for security teams: how to prevent employees and backend scripts from leaking proprietary business data to external JEV endpoints.
What Is JEV?
JEV is the first model in a class that TypeSafe AI calls "System One" models. The name comes from Kahneman's distinction between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). JEV is designed for fast decisions.
Traditional large language models (LLMs) generate an answer one token at a time. JEV does not work this way. According to TypeSafe, JEV takes unstructured program state as input. It returns a typed, probabilistic decision in one parallel pass. Each output includes a confidence score.
What JEV Does and Does Not Do
| Feature | Traditional LLM | JEV (System One) |
|---|---|---|
| Output type | Free text, code, summaries | Typed decisions, classes, probabilities, scores |
| Generation method | Sequential token generation | Single parallel pass |
| Writes text | Yes | No |
| Response time (TypeSafe claim) | Seconds | 70 ms to 500 ms |
| Price (TypeSafe claim) | $0.20 to $10 per million tokens | Input: $0.042 per million tokens. Output: free |
| Type errors | Possible | Structurally impossible, according to TypeSafe |
- Classification: JEV assigns a record or event to one of the predefined classes.
- Routing: JEV decides the next step in a workflow.
- Scoring: JEV assigns a value to an input with a confidence score.
- Verification: JEV checks the output of another LLM.
- Real-time loops: JEV runs in environments where speed is critical, such as games, robots, and simulations.
Known Limits
JEV is currently in early access. TypeSafe states a limit of 255 classes for single-stage decisions. The model processes text and structured state for now. It does not process images. Speed and cost figures are the vendor's own claims. Do not treat them as final until you measure them independently.
Why JEV Matters for Enterprise Security
JEV does not generate text. Some teams can therefore assume it is low risk. This is wrong. The risk is in the input, not the output.
JEV exists to make decisions about business data. To classify a customer record, route a support ticket, or score a transaction, the model must read the raw data. That data goes to a service outside your company. This is the same type of data exit as pasting a customer list into a ChatGPT prompt.
- Shadow AI: A developer can register on the JEV console without IT approval and test with real data.
- Bulk data flow: JEV is recommended for map-reduce operations on large datasets. This means much more data than a single prompt.
- Automation and API use: Scripts and workflows can call JEV in the background. The user sees no warning.
- Data residency: TypeSafe states that its service is hosted on the US West Coast. For personal data, this raises the question of cross-border transfer under GDPR and similar laws.
- Early-access uncertainty: The product is new. Each organization must confirm the retention period and data processing terms with its legal team.
How Moftgate Controls JEV Access
Moftgate manages AI at the endpoint level. You do not need a separate tool for JEV. Your existing AI Governance and AI DLP controls also apply to JEV traffic. The control works in three steps.
1. Visibility: Find Out Who Uses JEV
The Moftgate agent maps running processes and active network sessions in real time. New AI services are detected automatically. You can see which device, which user, and which application connects to a JEV service. Browser, desktop application, and API client traffic are monitored separately.
2. Policy: Block, Warn, or Allow
Each organization has different JEV needs. Moftgate gives you three policy models:
| Policy | When to Use | Result |
|---|---|---|
| Full block | Legal or compliance has not yet approved JEV | Connections to JEV services stop at the device level |
| Allow with warning | The product is in the evaluation phase | The user sees a policy warning and the access is logged |
| Controlled allow | JEV is used in an approved workflow | Only authorized users and devices connect. AI DLP blocks or masks sensitive data |
For the block rule, add the TypeSafe console address (console.typesafe.ai) and the API addresses published in the TypeSafe documentation to your rule list. TypeSafe can publish new domain names, so check this list regularly.
3. AI DLP: Stop Sensitive Data Before It Reaches JEV
Even when JEV access is approved, not all data must be sent. Moftgate AI DLP inspects the content at the moment of sending.
- Sensitive data detection: Data such as national ID numbers, credit card numbers, API keys, and intellectual property is found with regular expressions and machine learning models.
- Block or mask: Critical data is anonymized before it is sent, or the operation stops completely.
- User alert: The user sees the policy violation immediately.
- Audit log: Each event is recorded for GDPR and other compliance audits.
Implementation Note
The JEV output is a structured decision, not text. Focus on input control, not output analysis. Stopping the data before it is sent removes the risk at its source.
Step by Step: Set Up a JEV Policy
- Start with visibility. In the first week, monitor JEV traffic without blocking. Identify who is testing it.
- Collect use cases. Ask users why they use JEV.
- Do a legal and compliance review. Evaluate the data processing terms, the retention period, and cross-border transfer.
- Choose the default policy. Apply full block or allow with warning until a decision is made.
- Define exceptions for approved use cases. Set an authorized user group and a device list.
- Enable AI DLP rules. Block sensitive data even in approved use.
- Review reports regularly. Update the policy as the early-access product changes.
Strategic Takeaway
Banning JEV is not the solution. The model can make fast, low-cost decisions, and this is real business value. The correct approach is to make usage visible and to protect sensitive data at its source.
Research team specializing in enterprise LLM security, AI governance, and model-level data loss prevention.
Expert Answers to Core Practitioner Questions
What is JEV?
JEV is a System One model developed by TypeSafe AI. It does not generate text. It reads structured input and returns a typed, probabilistic decision.
Is JEV a chat model?
No. JEV does not write replies, generate code, or produce summaries. If you need text, you must use a separate LLM.
If JEV does not generate text, why is it a data security risk?
The risk is in the input. To make a decision, JEV reads raw data such as customer records, transaction data, or document content. This data is sent to a service outside your company.
Can Moftgate block JEV access?
Yes. Moftgate detects connections to JEV services at the endpoint level. It blocks, warns, or logs them based on your policy. For approved use, AI DLP also stops sensitive data from being sent.
Does Moftgate see JEV traffic from API clients too?
Yes. Moftgate monitors browser prompts, desktop applications, and API clients at the endpoint level.
Are the JEV speed and cost figures accurate?
These figures are TypeSafe's own claims. The price of $0.042 per million input tokens and the response time of 70 ms to 500 ms come from the vendor. Do your own measurements before you use JEV in your organization.
Authoritative Sources & Regulatory Standards
- •TypeSafe AI: Introducing System One Models & Jev ↗(Official vendor announcement)
- •
- •
- •
- •