The model proposes. The host decides.
Host code outside the model answers what an agent asks: an approval gate, not an operating-system jail.
Read what it does not enforceWhere the boundary is
Two ways to read one boundary: who decides what runs, and where your data goes. Each label opens its row in the table under the map.
Who decides what runs: the model proposes, host code checks, you answer what it asks, and what you approve runs as you.
- The modelProposes a tool call. Nothing it says is a decision.
- Checked in codeNarrower than it sounds
- The host gateDeny-list, task mode and the run's grant, checked in code.
- Only you answerNarrower than it sounds
- YouAnswer what your approval mode sends you.
- Runs as youNarrower than it sounds
- What runsRuns with your user's reach, or in the container or VM tier you chose.
- An HMAC per entryNarrower than it sounds
- The recordThe security log, with an HMAC over each entry.
- Holds
- True at this release, with a live consumer rather than only a declaration.
- Narrower than it sounds
- Something real ships, but the obvious reading of the claim is wider than what this release does. The row states the narrower truth.
| Boundary | Responsibility | Enforcement | Evidence | Known limit |
|---|---|---|---|---|
| The model's proposal to the host gateAuthority | A tool call is only a proposal. Host code checks the deny-list, the task mode and the run's grant before any approval is consulted, so not even the most permissive approval mode can run what a task mode refuses. | Narrower than it sounds | agents/ | An agent CLI meets these checks only for the calls it asks about. |
| The host gate to youAuthority | Your approval mode decides which calls wait for you, and only you answer them: an agent's tool or an app cannot approve what waits on you. | Narrower than it sounds | approval_answer.py | Per-tool grants bind subagents, workflow leaves and unwatched turns, not the agent you chat with. |
| Your decision to what runsAuthority | An approved command runs with your user's filesystem and network reach, with the gateway's secrets left out of its environment. A container or virtual-machine tier is opt-in. | Narrower than it sounds | sandbox.py | The agent's shell is screened, not fenced, from your credential files. |
| What runs to the record of itAuthority | Each security event is written with an HMAC over its own record, so an entry edited in place fails the integrity check. | Narrower than it sounds | sel.py | A deleted entry goes unreported: the check allows gaps, because several processes write one log. |
| Your network to the gatewayNetwork | The gateway asks for a local token by default. The one mode without it binds to loopback, so turning auth off cannot open the gateway to your network. | Holds | auth/ | A process running as you can answer as you. |
| The gateway to an appNetwork | An app reaches the gateway with a token scoped to that app, through the gateway's own proxy, and an undeclared write is refused. | Narrower than it sounds | dashboard/ | An app's own code runs as you, and its backend runs on the host unless its manifest names a sandbox tier. |
| The gateway to the internetNetwork | Core's fetches, the model clients and an app's SDK requests pass one guard that refuses private, loopback and link-local addresses and dials the address it checked. | Narrower than it sounds | net/ | By default it is not a destination allow-list, and an app backend's own traffic does not pass it. |
| The gateway to the provider you choseNetwork | What a prompt is built from is masked before the model is handed it: a key, a token or a URL's login is found by its shape or its field's name. | Narrower than it sounds | security.py | A secret written as a plain word reaches the model as written, and an agent CLI's own tools are outside the mask. |
What it does not enforce
Read against v0.3.9, the last release we verified, on November 16, 2026. Each one links where it is stated in full.
Read every limitationNot an operating-system jail. By default an approved command runs with your user's filesystem and network reach; the container and VM tiers are opt-in.
Limitations: the shell is screened, not fencedAn agent CLI meets the host's checks only for the calls it asks about. What its own settings let it run without asking is shown and logged after it has run.
Limitations: an agent CLI's own callsAn app's backend is confined only when its manifest names a sandbox tier. By default it runs on the host, where its network permission is a disclosure, not a fence.
On Compare, narrower than it sounds: “Apps are sandboxed, with a per-app network policy”Per-tool grants hold subagents, workflow leaves, room members and unwatched turns, not the agent you chat with, and only the hand-edited ceiling file names tools.
On Compare, narrower than it sounds: “Per-tool grants restrict which tools an agent may call”The egress guard protects your network, not a list of destinations: by default an agent can reach any public host.
On Compare, narrower than it sounds: “Outbound network access is deny-by-default”An exclusive host list for agent work exists only through the hand-edited governance ceiling, read once at start. Without it, an Allowed hosts entry adds to the default.
On Compare, narrower than it sounds: “You can restrict outbound access to an explicit list of hosts”The security log catches an edited entry, not a deleted one.
On Compare, narrower than it sounds: “Autonomous actions leave an audit trail that would show tampering”The dependency audit stops a merge that adds a dependency with a known high or critical vulnerability, not one that turns out vulnerable after it merged.
On Compare, not in this release: “A vulnerable dependency cannot reach the default branch”
The controls, and the code that enforces each
Each links the file that enforces it at v0.3.9. A control is listed only if that code has a live caller at the tag; a setting nothing reads is not a control.
Task modes
A tool the task mode does not allow is refused before any approval is asked, so an auto-approve cannot run it.
agents/native/ _guard_and_invokeruntime.py , Approval modes
Your approval mode decides which tool calls wait for you. Auto-approve expires, and turning it off fires every surface's disable.
trust_mode.pyOnly you answer
An ask that waits on you is answered by you: a request an agent's tool or an app sends cannot approve it.
approval_answer.py, refusalCommand deny-list
Known-destructive shell commands are refused whatever the mode or the approval. The list reads a command's text.
shell_denylist.py, BUILTIN_DENIED_COMMAND_PATTERNSGuarded egress
Requests out are refused for private, loopback and link-local addresses and dial the address that was checked. By default it is not a destination allow-list.
net/guard.py evaluate, Fenced outside content
A fetched page and other outside text reach the model marked as data, not as instructions.
security.py, fence_untrustedA token by default
The gateway asks for a local token, and the one mode without it binds to loopback only.
auth/modes.py effective_bind, Optional sign-in
A password, argon2id-hashed, and a second factor, for reaching the dashboard from another device. Off by default and additive: the local token keeps working, so a forgotten password cannot lock you out.
dashboard/handlers/ the login doorauth.py , Scoped app tokens
An app calls the gateway with a token scoped to it, through the gateway's proxy. Its own code still runs as you, and its install review says so.
dashboard/token_auth.py app_session_token, Scanner gate
An app is scanned in quarantine before you consent to it, and a dangerous verdict cannot be overridden into an install.
apps/app_manager.py install, Tamper-evident entries
Each security event carries an HMAC over its own record, so an entry edited in place is caught. A deleted entry is not.
sel.py, verify_integrity
An app is code, not a collectible.
Installing one is a staged trust decision. The package stays apart until it is scanned and you have consented to what it declares.
supply_chain.py, Verdict- QuarantineThe package is staged apart, inactive, and never follows a link out of its bundle.
- ScanThe scanner reads the staged tree, which is the tree that would be installed, and gives a verdict.
- ConsentYou see what it declares, including what the platform does not enforce, before anything is enabled.
- EnableOnly then is it installed, with access the gateway issues. A dangerous verdict ends it at the scan.
Report a vulnerability
Core's security policy says how to report one privately and what to expect back. The threat model behind this page names each trust boundary and maps the OWASP agentic risks to the code that answers them.