Security
CommandGovernor
The CommandGovernor sits between LLM tool decisions and shell execution. It runs before run_command, run_script, and manage_process.
Blocked (CRITICAL)
These patterns are blocked before the command reaches a shell. With tools.governor.admin_can_override: true (the schema default and the tracked template), an admin-tier requester can override the block; the override is annotated in the output and recorded in the audit log. Set it to false to make the block absolute.
| Pattern | Reason |
|---|---|
rm -rf / | Recursive delete on root |
mkfs | Filesystem format |
dd if= | Raw disk write |
| Fork bombs | System destabilization |
shutdown, reboot, halt | System control |
chmod -R 777 / | World-writable root |
iptables -F | Firewall flush |
ufw disable | Firewall disable |
DROP TABLE, TRUNCATE | Database destruction |
crontab -r | Remove all cron jobs |
> /dev/sd* | Block device write |
Always Blocked: Unconditional Git Force Pushes
The governor blocks literal git push invocations that use --force, -f (including short-option bundles), or a force-prefixed +<refspec>. This rule is not admin-overridable. --force-with-lease and --force-if-includes are not mistaken for unconditional force, but combining either with --force, -f, or a + refspec is still blocked.
This is a bounded, quote-aware recognizer for literal shell commands. It handles ordinary shell control operators, Git global options such as -C, environment assignments, and common sudo, env, command, exec, and nohup wrappers. The literal tokenizer preserves comment-ending newlines, joins unquoted and double-quoted backslash-newline continuations, and distinguishes quoted or escaped separator arguments from actual operators. A mid-word # is literal. It recognizes else branches and separate or attached values for env -u / --unset and -C / --chdir. It is not a complete shell interpreter: aliases, functions, runtime expansion, eval, generated arguments, and command strings handed to another interpreter cannot be resolved statically. Redirections, heredocs, compound shell grammar, and interpreter-specific syntax are not fully parsed. An allowed result is not proof that arbitrary shell code cannot force-push. Those limitations must not be treated as a safe way to bypass the policy.
This rule does not replace the removed git_ops push safeguards. It does not perform the former freshness preflight or stale-branch refusal, bind the source and destination to the preflighted commit, obtain an observed remote SHA for an exact lease, or refuse a push when that evidence is missing. Operators using shell Git commands must do those steps explicitly before an allowed --force-with-lease=<destination>:<observed-sha> push.
Blocked (Exfiltration)
| Pattern | Reason |
|---|---|
curl | bash, wget | sh | Pipe remote script to shell |
bash -i >& /dev/tcp/ | Reverse shell |
nc -e /bin/sh | Netcat reverse shell |
python -c socket.connect | Python reverse shell |
base64 -d | bash | Obfuscated payload |
> /etc/passwd, > /etc/shadow | Auth file writes |
>> /etc/cron* | Cron persistence |
Allowed with Annotation (HIGH)
Commands classified as HIGH risk are allowed but annotated in tool output:
[governor: allowed — high risk, recursive delete]Examples: rm -rf /tmp/specific_dir, systemctl restart nginx, and docker rm container.
Denial Response
When a command is blocked, the tool returns a structured denial:
Blocked [critical]: recursive delete on root
Suggested alternative: Use a more specific path, e.g. rm -rf /tmp/specific_dirPermissions
Three tiers control tool access:
| Tier | Tools | Use Case |
|---|---|---|
admin | All enabled built-in tools | Operators |
user | 11 tools: ten read-only plus list management | Team members |
guest | None (chat only) | Restricted |
User-tier tools: get_tool_output, search_history, search_knowledge, web_search, fetch_url, list_schedules, list_tasks, list_skills, list_knowledge, manage_list, parse_time. run_command is deliberately excluded (arbitrary shell execution). manage_list can change list state; the rest are read-only. Host access is enforced separately per user at execution time.
Set default in config, override per-user via set_permission or web UI.
Secret Scrubbing
All input/output paths scrub secrets:
- Discord messages (inbound) — detected and deleted
- Tool results (outbound) — scrubbed before display
- Audit logs — tool_input string values scrubbed
- Bot message buffer — scrubbed before entering LLM context
- File attachments — scrubbed before context injection
Patterns detected: Discord tokens, API keys, Bearer tokens, SSH private keys, AWS credentials, database URLs with passwords.
Web API Authentication
- All
/api/*endpoints require Bearer token auth whenweb.api_tokenis configured - Session-based auth via
/api/auth/login - Web chat uses server-side identity — caller cannot spoof
user_id - WebSocket connections authenticate without placing credentials in the URL. Browser clients offer
odin.bearer.<base64url(token, unpadded)>as a WebSocket subprotocol; non-browser clients may send anAuthorization: Bearer <token>header./api/ws?token=...is explicitly rejected (close code 4001). Do not put credentials in query strings or logs.
Chat attachments
A WebUI or REST chat question can carry files. The page uploads each file first (POST /api/chat/uploads, multipart, one file per request) and names the returned ids in the question's attachments. The files then go through the same processing as a Discord message's attachments, under the same attachments.* limits.
A staged file belongs to the identity that uploaded it. Another identity's id is refused, and a file is used by one question only.
Limits on this upload path:
- 50 MiB per file and 10 files per question;
- at most 100 files and 500 MiB staged per identity, and 1,000 files and 1 GiB in all, counting uploads still streaming;
- one upload request may take up to 15 minutes.
A full store refuses new uploads. It never deletes a staged file that hasn't expired.
Staged files live in
data/web-uploads/: the directory is 0700, the files are 0600, and the server chooses their names (a client's filename is only a label).A question's files are deleted once processed, and unused ones after 24 hours. After a restart, files a question had claimed are deleted, never used again.
A claim that can't be recorded (a storage error) refuses the question. Its files go back to staged, to send again, or are deleted if even that can't be recorded. An upload request that ends without returning its file's id (a deadline, a malformed or extra part, a dropped connection) keeps nothing.
Back up the data directory without
data/web-uploads/: it holds only transient uploads./api/executetakes no attachments.
Skill Sandboxing
Skills run in-process as trusted plugins (not sandboxed). Mitigations:
- Admin-only creation by default
- Validation versus loading — standalone AST validation does not execute code. Trusted local create/edit loads Python before validating the runtime definition; URL install additionally runs static validation and prohibited-construct checks before loading. None of these checks is a sandbox.
- Safe tool allowlist — skills cannot call
run_command,apply_patch, etc. directly - Blocked file paths — cannot read
.env, SSH keys, credentials - Blocked URLs — localhost, private IPs blocked unless whitelisted
- Resource limits — 120s timeout, 50 tool calls, 20 HTTP requests
Codex Auth
- OAuth tokens stored with 0600 permissions
- Atomic writes (temp + fsync + rename) prevent corruption
- Existing split-account files preserved on restart (no stale overwrite)
- HTTP requests to token endpoint have 30s timeout