An agent on a computer is a serious thing

Software that can see a screen and run commands deserves more scrutiny than a web app. This page describes current practice, honestly and without guarantees — it changes as afire matures, and the changelog records when it does.

Where afire is today

• Closed alpha. Access is by invite and capacity is deliberately small. There are no formal certifications yet (SOC 2, ISO 27001); security questionnaires are answered on request.
• Choose devices accordingly. Alpha means rough edges. Connect devices you’d be comfortable handing to early software — not the one system a business can't lose.

Practices

• Signed updates. Agent releases are signed, and each agent verifies the signature on-device before applying an update. An unsigned or altered release is refused. Signing keys are kept off the production environment.
• Outbound-only agents. Devices connect out over TLS. The agent doesn't listen for inbound connections on your hardware.
• Reduced privileges. The platform service runs as an unprivileged user, and agents only receive tools the device supports and the session allows — a headless server doesn’t expose screen control.

Accountability

• Actions are logged. Commands, input control, screenshots, and file operations are recorded with their source and time, reviewable per device.
• Two-factor auth. TOTP on accounts; administrative actions are audited separately.

Data

• Deletable. Deleting a device deletes its chat history and stored media. Chat attachments are stored unopened server-side and expire after use.
• Diagnostics are opt-out. Sharing diagnostic data is a setting, not a condition.
• Metered with caps. Usage runs against per-account daily and monthly ceilings, enforced server-side.