Rabbit Chat Studio tools have real system permissions. The goal of the boundary design is not "make the model absolutely safe", but to make what it can do, what it did, and why it was rejected visible.
Threat model
The main risks are not "a malicious model" but:
- misjudged paths deleting data outside the workspace or inside
.data; - prompt injection (web pages, files, MCP responses) inducing dangerous commands;
- auto-accept being left on, so dangerous commands run unchecked;
- third-party MCP servers executing code with your permissions;
- API keys, database passwords and business data ending up in logs or exports;
- the assistant rewriting its own instructions in self-evolution mode.
Path boundaries
The three execution tools have deliberately different scopes:
| Tool | Where it may reach | How it is bounded |
|---|---|---|
file_actions |
The workspace root, including the hidden .data folder |
Path sandbox — no escape hatch |
run_cmd |
One PATH-resolved program | Name check + risk assessment |
run_shell |
Anywhere, including outside the workspace | Risk assessment + confirmation |
file_actions
- can only operate inside the workspace root;
- accepts
./...,., and absolute paths inside the workspace; - absolute paths outside the workspace,
../escapes and rooted paths (/etc,C:\) are rejected; - canonicalizes paths and checks symlinks;
- cannot write/delete/rename the workspace root itself;
.datais part of the workspace — it is listed, searched, read and written like any other folder, so the assistant can look up and manage its own skills, memories and todo lists;- the one exception is the workspace database (
.data/chat.dband its-wal/-shm/-journalsidecars):file_actionscould only truncate or delete it, never rewrite it, so it is rejected — it holds the chat history, the request monitor, the interaction log and the agent missions; - the one skills root (
./.data/skills) cannot be deleted as a whole; - everything under the skills root gets a
.bak.Nbackup before it is overwritten or deleted.
run_cmd / run_shell
- processes start in the workspace root;
run_cmdruns a bare program name that the OS resolves throughPATH(python,python3,git, …); absolute and relative paths are rejected before the risk assessment runs;run_shellscripts cancdelsewhere; the tool does not prevent it;- path boundaries for the shell rely on risk assessment and confirmation, not a sandbox;
- use
run_shellfor external paths and expect higher risk levels.
MCP
- stdio MCP cwd is the workspace, but it is not a sandbox; it can access anything your user can;
- remote MCP servers receive the arguments you send;
- MCP tool arguments are treated as opaque; the app does not path-check them.
CLI / cloud sub-agents
- CLI agents run with your user permissions;
cwdis sandbox-checked as a workspace-relative directory, but the process itself can still access other paths; - cloud agents send the task to a remote endpoint; data leaves the machine;
- both require approval on first run by default (
external_agent).
Tool permissions
Three layers stack:
- Global tool switches (
config.selected_tools): which built-in tools the main agent sees; - Skill command allowlist (
allowed-commands):run_cmdexecutable names must be in the union of enabled/loaded skills; - Sub-agent tool allowlist (
allowed_tools):llmsub-agents only get a subset offile_actions/run_cmd/run_shell.
Notes:
allowed-commandsis currently enforced only onrun_cmd;run_shellstill goes through risk confirmation;- sub-agents' panel only offers the three execution tools;
timeris always filtered; - MCP tools are not controlled by
selected_tools; only the server enable switch controls them; - background execution (
config.background_tools) does not add a new permission: it only lets an already-enabled tool run asynchronously, and the same risk assessment, confirmation, path checks and allowlists apply before the task starts; - self-evolution mode overrides the first layer:
file_actionsandmemoryare forced on and cannot be switched off, by the panel, a profile, a backup restore or a hand-editedconfig.json.
Risk levels and confirmations
run_cmd / run_shell run a static assessment:
| Level | Score | Behavior |
|---|---|---|
| L0 | ≥85 | Must confirm (dangerous) |
| L1 | 70–84 | Must confirm (system_config) |
| L2 | 55–69 | Must confirm (user_software) |
| L3 | 40–54 | Must confirm (user_data) |
| L4 | 25–39 | Runs directly and is logged |
| L5 | 10–24 | Runs directly |
| L6 | 0–9 | Runs directly |
The dialog shows the level, score, matched rules and penalty items. Waiting more than 120 seconds counts as a denial. A script that fails the syntax gate is rejected before it is scored at all.
Auto-accept risk
Tools → Auto-accept mode can skip L0–L3 individually, or everything via "*". When on:
- dangerous commands no longer ask;
- path checks, skill allowlists, risk assessment and logging still happen;
- but nothing stops an 85+ command anymore.
Recommendations:
- keep it off day to day;
- only enable it in an isolated workspace, briefly, for a clear task;
- never on a production machine, in a directory with secrets, or with unrecoverable data;
- turn it off immediately afterwards and inspect the interaction log.
Skill isolation
The skill root is the folder containing SKILL.md:
skill_readcan only read inside that folder;../escapes are rejected;- paths in the body are relative to the skill root;
- there is exactly one skill root per workspace,
<workspace>/.data/skills— it is computed in a single place in the backend, so a second location cannot be introduced; - Self-evolution mode lets the assistant write, extend and delete skills under that root, and creates
.bak.Nbackups for everything it touches.
Prompt injection
Skills, MCP responses, web pages and file contents can contain malicious instructions. The model may treat them as user requests. Reduce the risk:
- do not give untrusted content high-privilege tools;
- use least-privilege accounts for MCP servers;
- state in the system prompt that external content is not an instruction;
- keep human confirmation on for important operations;
- regularly inspect tool calls in the interaction log.
Process cleanup
- stdio MCP, serve and CLI sub-agents are registered in
process_guard; - app exit and window close kill whole process trees;
- Windows uses a job object, so even an app crash does not leave children behind;
- Unix uses process groups so
kill(-pid)reaches grandchildren; - serve processes are killed at the end of the current chat turn by design;
- background tool tasks are turn-scoped too: Stop or turn end sets every cancel flag and each runner kills its process tree;
- CLI agent timeout/Stop kills the process tree.
Secrets and data
api_keyinconfig.jsonis plaintext;auth_tokenandenvinmcp_servers.jsonare plaintext;interaction_loginchat.dbmay contain full request bodies, file contents and key fragments;- exported zips contain all of the above.
Recommendations:
- use full-disk encryption;
- do not put the app data directory in Git;
- add
.data/to the workspace.gitignore; - clear/compact logs periodically (Interaction Monitor);
- inspect zips before sharing;
- rotate leaked keys;
- use dedicated API keys and read-only database accounts in production.
Recommended security baseline
- [ ]
config.workspace_dirpoints at a dedicated project directory, not the whole disk or your home directory. - [ ] The workspace
.gitignorecontains.data/. - [ ] Only the tools needed for the current task are enabled.
- [ ]
run_shellis enabled only when pipes/builds are needed. - [ ] Auto-accept is off; if needed, use the smallest subset of L0–L3.
- [ ] MCP servers come from trusted sources and use least-privilege tokens.
- [ ] CLI/cloud sub-agents keep Require approval on.
- [ ] Production secrets are not placed in workspace files.
- [ ] Inspect
interaction_logand exported zips regularly.
Known limitations
run_shellcwd is not restricted to the workspace; scripts cancdout.file_actionscan write into.data(skills, memory, todos, skill status); only the workspace database is protected.- MCP and CLI/cloud agents are not inside the workspace sandbox.
allowed-commandsis enforced only onrun_cmd.- Risk assessment is a static rule set, not a runtime sandbox; obfuscated commands may evade the score.
- Auto-accept
"*"also covers future confirmation kinds. - Release builds do not write
app-<date>.log; use the Interaction Monitor and MCP logs. config.json/mcp_servers.jsonsecrets are not stored in the OS keychain.
For stronger isolation, run Rabbit Chat Studio in a VM/container, mount only the project directory, and keep MCP/CLI in the same isolated environment.