Security & privacy

Documentation Index

Fetch the complete documentation index at: https://docs.hypercubic.ai/llms.txt

Use this file to discover all available pages before exploring further.

Mainframe credentials are encrypted via the OS keychain, tap data stays in your user data folder, and the AI agent only sees your active prompt, recent chat context, and tool results.

​

Where data lives

​

On your laptop

Exact file paths live under your user data folder (macOS: ~/Library/Application Support/Hopper/, Windows: %APPDATA%\Hopper\, Linux: ~/.config/Hopper/).

​

On the mainframe

Hopper installs nothing on the mainframe. It connects over standard TN3270, plus FTP if you’ve added credentials. When you create a data tap (a VSAM snapshot), Hopper allocates a temporary sequential dataset under your HLQ to stage the data. Ask the agent to drop the tap when you’re done. If you don’t, it persists, but only within your HLQ permissions.

​

In Anthropic’s systems

When you use the agent, the API call sends your prompt and recent chat context to api.anthropic.com. Anthropic’s data policy governs what they do with that data. Their default does not train on API customer data, but check their current policy. What gets sent:

• Your active prompt
• Recent chat context (with hidden-field placeholders, no real values)
• Tool call results the agent needs to interpret
• A small system prompt with general mainframe context

What does not get sent:

• Mainframe credentials
• Hidden-field content from the terminal
• Anthropic API key (the auth header carries it, not the message body)
• Logs, tap data, or anything outside the active conversation

​

On Hypercubic-controlled servers

Nothing during normal operation. Hopper does not call back to a Hypercubic-controlled service except for update checks.

​

Credential handling

​

Mainframe credentials

Stored in the OS keychain. Hopper retrieves them when needed. They never enter the chat agent’s context. For actions that need credentials (e.g., a CICS signon), the agent calls a tool like cics_signon with no credentials in the call. Hopper’s sidecar process loads the credentials directly from the secure store and applies them. The model never sees the values.

​

Anthropic API key

Stored in the OS keychain. Used only to call api.anthropic.com. Never sent to anyone else, never logged.

​

Other secrets

If you have additional credentials (DB2 users, third-party API keys), Hopper does not know about them. They live wherever your mainframe applications expect them. Hopper does not intercept or store them.

​

Redaction

​

Hidden terminal fields

z/OS terminals expose hidden fields (typed input that’s not displayed, like passwords). Hopper masks the typed characters with bullets (•), one bullet per character, before they reach any boundary:

• Chat agent input: bullets only
• Logs: same

The mask runs inside Hopper. A UI bug cannot leak a hidden-field value.

​

Pasted secrets

If you paste secrets into chat, Hopper does not redact them. It cannot tell what’s secret. Treat the chat input the way you’d treat a Slack message. If you’ve pasted something sensitive by accident, click the trash icon at the top of the chat panel to clear the conversation.

​

Production data

​

Tap snapshots

A data tap extracts production records into a local file on your laptop. The file is not encrypted at rest beyond your filesystem’s normal protections. For regulated data (PII, PCI, PHI):

• Drop taps when you’re done.
• Use full-disk encryption (FileVault, BitLocker, LUKS).
• Don’t sync the tap file to cloud storage.
• Treat exported reports as having the same sensitivity.

If your organization prohibits extracting production data to laptops, don’t use taps on production data. Hopper has no way to enforce that policy from your side.

​

Reads

Reading a member or copybook does not persist a copy. The content lives in the agent’s chat context for that turn only. Logs may record the dataset name but not its content.

​

Logs

Hopper writes logs to your user data folder. They contain:

• Application events (startup, connection state, agent turn boundaries)
• Errors and stack traces
• Chat transcripts: your prompts, the agent’s responses, and tool calls. Tool inputs are truncated to 500 characters and tool results are capped at 30 lines, so full file content doesn’t land in logs — but dataset names, job IDs, and prompt summaries do.

They do not contain:

• Credentials
• Hidden-field content
• Full datasets or file content

Logs live in ~/Library/Application Support/Hopper/logs/ (macOS), %APPDATA%\Hopper\logs\ (Windows), or ~/.config/Hopper/logs/ (Linux). If you send logs to debug an issue, review them first — chat transcripts can include prompts and dataset names you’d rather not share.

​

Network

Hopper makes outbound connections to: Hopper makes no other outbound connections. It opens no inbound network listeners.

​

Telemetry

Hopper does not ship telemetry. Nothing about your usage, prompts, or mainframe activity is sent to Hypercubic.

​

Compliance

Hopper is not currently SOC 2, ISO 27001, or HIPAA certified. Hopper does not transmit your data to our servers, so the audit surface for Hopper itself is small. The relevant audit usually concerns your laptop, your network, and Anthropic’s API endpoint. For specific compliance questions, email team@hypercubic.ai.

​

Reporting a security issue

Email team@hypercubic.ai. We respond within two business days. We especially want to hear about:

• Anything that exposes credentials beyond the OS keychain.
• Anything that bypasses the approval gate for mutating actions.
• Network behavior beyond what this page describes.

​

Where to read next