Skip to main content
Axosun is currently unavailable, and its return to service is not planned in the near future.
The key limitations of Axosun are implemented in the product code, not in the model instructions, and the model cannot be “persuaded” to bypass them: the separation of read-only and destructive actions, authorised lists and mandatory confirmations are checked by the panel at every step.
For whom: Administrator/Security Specialist Required Rights: Viewing the Audit – Access to Operations Journal.

Fixed action catalogue

The assistant can only suggest actions from the product catalog. Unknown model sentences are discarded in the code before being shown to the operator. run_shell_command, but it does not give the assistant constant access to the shell: the command is run only for a specific node and only after a separate decision of the operator. Opportunities.

Mandatory evidence

Changing the firewall and restarting the server (including mass) always require confirmation — these actions cannot be removed from the list of confirmed by any configuration. sysctl And packet installation is completely prohibited. The model of rights.

Preliminary verification of requests

Another do reference model local classifier blocks requests for:
  • disclosure of secrets;
  • Disclosure of hidden instructions of the assistant;
  • commands without a specific node or outside the action of run_shell_command;
  • cloning of the product and mass unloading of the knowledge base;
  • Messages with signs of prompt injection
Such requests are rejected without a single reference to the model (and without the consumption of tokens), and the refusal is recorded in the audit — with a short fragment of the message, not with the full text.

Protecting team launches

Before showing the card, Axosun rejects an empty command and obviously disastrous templates. In particular, you can not approve the removal of the root or system directory through rm -rf, formatting through mkfs, writing dd to a block device, erasing the partition, turning off or rebooting the host and fork bomb. The card contains the exact command, the node, the cause, the level of risk and the absence of automatic rollback. Once confirmed, Axosun rechecks the dangerous patterns before setting the task to the agent. The confirmation of the primary protection is not replaced: it is mandatory for each team. Blocking intentionally covers only obviously disruptive options; it does not make an arbitrary command secure, nor does it remove the operator’s obligation to verify its meaning, purpose, and consequences.

Protection against looping

Tool calls are traced (tool + goal + arguments) to recognize the repetition of the same work. If a destructive action for the same purpose fails twice without changing state, further attempts are blocked, the incident escalates to operator status, and the assistant is instructed to stop repeating. Failed actions are never restarted automatically.

Launch budgets

Agent launch is limited to 8 steps and 6 minutes; after exhaustion, a report is stored marked with the reached limit. See Sessions and launches.

Reporting discipline

After analyzing the model response and before saving the report, deterministic post-processing is applied:
  • If the system is demonstrably healthy, all recommendations are removed and the canonical conclusion “no action is required” is added.
  • report_only mode is guaranteed to contain no destructive steps - if the model still offers such steps, they are deleted in the code;
  • The completed start-up of direct action or repair does not carry the “tails” of additional offers.

Mandatory validation of actions

  • Each deterministic direct action is accompanied by a mandatory read-only check of the result (for example, after a firewall change, a security profile check; after a WARP restart, a WARP status check). If the check fails, the launch is recorded as unsuccessful.
  • In the classic mode, after a successful destructive action, the panel automatically performs a read-only check “fix worked” and saves the result in the report.

Incident life cycle assurances

  • Incident statuses change only on permissible crossings; an attempt at an inadmissible crossing is rejected.
  • Terminal statuses are “sticky” – it is impossible to accidentally overwrite a closed incident (only an explicit re-opening is allowed).
  • Repeated reports update descriptions, but do not reset the status set by the operator.
  • “Decide in order” only builds the sequence of open incidents – no corrections are performed automatically.

Diagnostic configuration invariants

In a Flow diagnostic scenario, configuration validation is performed first and unconditionally; non-valid configuration takes precedence over discrepancy (drift) when classifying an incident.

Full audit.

The log of operations records:
  • Axosun policy changes;
  • Creation/modification/deletion of model profiles
  • the life cycle of sessions and launches;
  • Every confirmed and rejected action;
  • each call of the Remnawave tool - with sanitized arguments and an impersonal snapshot of settings;
  • Failures of preliminary verification of requests.
The audit_log notification channel does not turn off: the panel automatically returns it to the policy when saved.

Verification

1

Check the audit.

Perform a trial run and confirm or reject any action, then open the Operations log - the launch and action decision records must be present.
2

Check pre-filtering.

Ask the assistant to show you the secrets: the request must be rejected immediately. Then request a harmless verification on a specific node, such as df -h: a card with an exact command should appear, not an immediate launch.
Actions and failures are visible in the audit; requests for disclosure of secrets are blocked, and any team waits for confirmation.

Rollback

The fuses described on this page do not turn off - these are the built-in properties of the product. The customized part of the restrictions (behavior modes, access to Remnawave) is described in the Rights Model; there is also how to return the default values.