AI server manager authority modes

The TaraSec AI server manager observes an Ubuntu server, asks an OpenAI model for a structured assessment, and applies only actions permitted by local policy. The server owner selects the authority level in /etc/tarasec-server-manager.conf.

Default: New installations use AI_AGENT_MODE=conservative. The manager cannot change its own mode.

Accepted values

Value	Intended authority	Approval policy
`disabled`	No AI assessment or AI-initiated action. Conventional TaraSec status reporting may continue separately.	No actions can be approved or applied.
`conservative`	Diagnose, collect bounded evidence, notify and propose actions.	A local human must approve every system-changing action.
`defensive`	Intended for automatic, reversible containment of an active threat, with immediate reporting and verification.	High-impact or persistent changes still require approval.
`autonomous`	Intended for broad local investigation and repair on an intentionally delegated server.	Destructive, evidence-erasing or management-access-threatening actions remain protected.

Configuration

# disabled, conservative, defensive, or autonomous
AI_AGENT_MODE=conservative

# Services available to the current pilot
SERVICE_ALLOWLIST=taralink.service,tarasec-ssh-honeypot.service

OPENAI_MODEL=gpt-6-astra
STATE_DIRECTORY=/var/lib/tarasec-server-manager

After editing the file, start a new assessment. The manager rereads the mode before approval and again immediately before execution. Unknown or misspelled values generate a warning and use conservative; they never increase authority.

Current implementation status

Pilot limitation: The current draft implements structured diagnosis plus none or a human-approved restart of a service in SERVICE_ALLOWLIST. Selecting defensive or autonomous does not yet give the agent additional commands.

The broader modes define the policy direction for the OpenAI agent work. Planned capabilities include richer CPU, load, memory, swap, filesystem and service monitoring; operator notification; TaraSec requests for assistance; additional diagnostics; reversible containment; and post-action verification. Each capability must be implemented, audited and tested before the mode can use it.

Safeguards in every mode

Every assessment and action records the active mode and timestamp.
Logs and diagnostic output are treated as untrusted evidence, not instructions.
Actions and results are written to a local audit trail.
Temporary containment should expire or be explicitly released.
The agent must preserve evidence and verify the result of an action.
The local configuration remains authoritative; the model cannot raise its own authority.