diff --git a/README.md b/README.md
index 3948cac..a31f0fb 100644
--- a/README.md
+++ b/README.md
@@ -150,7 +150,8 @@ sandboxed agent never gets `/mcp`; it gets a **reverse attach** with the `deskto
 
 ## Lending this Mac to a sandboxed agent (reverse attach)
 
-An agent in an `openab-pty` session cannot reach this Mac — the pod has no egress by design. So
+An agent in an `openab-pty` session cannot reach this Mac — the pod has no tailnet egress by design
+(it does reach the internet, for its model API, git and package registries). So
 **this Mac dials the pod** and serves MCP over that socket with a narrowed tool list. Design:
 [`docs/adr/reverse-attach.md`](docs/adr/reverse-attach.md); wire contract: openab-pty
 `CLIENT-CONTRACT.md` §9.
diff --git a/docs/adr/reverse-attach.md b/docs/adr/reverse-attach.md
index ef80ed6..66710ba 100644
--- a/docs/adr/reverse-attach.md
+++ b/docs/adr/reverse-attach.md
@@ -15,7 +15,9 @@ Connect.
 Three facts about the sandbox constrain the answer:
 
 1. The shell runs as `uid 1000` with no host credentials and a read-only rootfs. openab-pty's
-   standing invariant is that **the shell never holds a credential**.
+   standing invariant is that **the shell holds no host or cloud credential**. (It is not
+   credential-free: the agent CLI's own model login lives in the container, and a session may
+   carry others, such as git.)
 2. The tailscale sidecar is the pod's only tailnet identity and it is **inbound-only**:
    tailnet → sidecar → loopback runtime, via `tailscale serve`.
 3. The sidecar runs `--tun=userspace-networking` on both k8s and ECS. There is no tun device,
@@ -38,7 +40,12 @@ why we start from the opposite premise.
 
 ## Decision
 
-**The sandbox gets no egress at all. The Mac dials the pod.**
+**The sandbox gets no tailnet egress. The Mac dials the pod.**
+
+"Egress" in this ADR means **tailnet** egress. The pod's ordinary internet egress is open by
+default — the agent CLI needs it for its model API, git remotes and package registries, and the
+ECS tasks run with `assignPublicIp: true`. Restricting it is a deployment choice (an egress
+allowlist), not something this design provides.
 
 1. **openab-pty runtime** gains an inbound endpoint `WS /tools/attach` on its existing loopback
    listener, reached through the sidecar's existing `tailscale serve` exactly like
@@ -58,7 +65,8 @@ why we start from the opposite premise.
    `key` and `mouse` each reach the desktop user's shell, so `desktop` is shell-equivalent and the
    earlier rationale — "no `exec` because the agent already has a shell" — described a
    convenience, not a restriction. The profile was renamed to say so; the pod-side isolation (no
-   egress, credential-less shell) is unaffected and remains the boundary this ADR claims. A real
+   tailnet egress — internet egress is open by default — and a shell without host credentials)
+   is unaffected and remains the boundary this ADR claims. A real
    narrower tier must be built from tools that cannot reach a shell (`observe`: `sys_info` +
    `screenshot`; `browser`: Playwright with a per-grant throwaway profile), and the hands
    themselves are isolated only by lending a dedicated machine or VM.
diff --git a/docs/requirements/reverse-attach.md b/docs/requirements/reverse-attach.md
index 74c576f..7bf5e4a 100644
--- a/docs/requirements/reverse-attach.md
+++ b/docs/requirements/reverse-attach.md
@@ -1,4 +1,4 @@
-# Requirement — reverse attach: macmini dials the sandbox, the sandbox has no egress
+# Requirement — reverse attach: macmini dials the sandbox, the sandbox has no tailnet egress
 
 > Recorded 2026-09-24. Implements **Phase 4** of
 > [`connect-closed-loop.md`](connect-closed-loop.md): "Connect's agent (in the `openab-pty`
@@ -14,8 +14,8 @@ key / osascript / exec) so the agent can drive the Mac while the human watches i
 The sandbox shell has **no network identity of its own** (`uid 1000`, no host creds, read-only
 rootfs); the tailscale sidecar is the pod's only tailnet identity and it is **inbound-only**
 (tailnet → sidecar → loopback runtime). Every earlier idea started from "give the sandbox an
-egress path". This document starts from the opposite premise: **the sandbox gets no egress at
-all — macmini connects *in*.**
+egress path". This document starts from the opposite premise: **the sandbox gets no tailnet
+egress — macmini connects *in*.** (Internet egress is a separate matter and is open by default.)
 
 ## Rejected — every egress-based variant shares one root flaw
 
@@ -66,7 +66,7 @@ on the instance-mcp side.
   │                  │ (sha256 verifier)│    │ inbound  │  └────────────────────────┘  │
   │                  └──────────────────┘    │  only    └──────────────────────────────┘
   │   sidecar: inbound only — UNCHANGED      │                       ▲
-  │   no egress of any kind                  │                       │ "lend my Mac to this
+  │   no tailnet egress (internet: default)  │                       │ "lend my Mac to this
   └──────────────────────────────────────────┘                       │  session, 1h" — human,
                                                                      │  in OpenAB Connect
 ```
@@ -196,8 +196,8 @@ remainder of the grant TTL.
 |---|---|
 | sandbox needs a tailnet egress path | **none needed**; sidecar unchanged |
 | credential in the pod that a compromised shell could steal | **none** — pod holds a sha256 verifier only; a verifier cannot be used to connect anywhere |
-| compromised agent reaches other tailnet nodes | **impossible** — no egress to bypass through |
-| compromised agent uses macmini's full shell via `exec` | **blocked by profile** — the `sandbox` profile omits `exec`, or requires a human tap in Connect (the human is already watching the screen) |
+| compromised agent reaches other tailnet nodes | **no path**, provided the k8s node itself does not run `tailscaled` (otherwise apply openab-pty `networkpolicy-no-tailnet-egress.yaml`). The pod still reaches the internet by default |
+| compromised agent uses macmini's full shell via `exec` | **not blocked by removing `exec`.** Superseded (#45): GUI control is a shell, so `desktop` (formerly `sandbox`) is shell-equivalent; only `observe` is a boundary. See `docs/tool-profiles.md` |
 | compromised agent calls the tools it *was* granted | **residual, by design** — giving an agent hands carries this in every design; the difference is the hands are exactly as large as the human chose, for as long as they chose |
 | macmini's attach credential leaks | attacker could serve a fake `/tools/attach`? No — the credential authenticates macmini *to the pod*; a leaked one lets an attacker impersonate macmini to that pod, not reach macmini. Mint per-attach, TTL'd, revoke by deleting the verifier |
 | **the k8s node itself is a tailnet member** (homelab k3s boxes often are) | **out of scope, must be documented as an assumption** — pods ride the node's tailscale routing (`ip rule … lookup 52` → `tailscale0`, masqueraded as the node) and reach every tailnet peer regardless of the sidecar. Measured on p1 in 4a: with the host's tailscaled running the shell reached macmini/black over their tailnet IPs; with it stopped, no path at all. Deployment rule: **no tailscaled on the node**; if unavoidable, an egress `NetworkPolicy` denying `100.64.0.0/10` + `fd7a:115c:a1e0::/48`. Fargate has no such path |
diff --git a/docs/tool-profiles.md b/docs/tool-profiles.md
index 5d0344e..ececf4b 100644
--- a/docs/tool-profiles.md
+++ b/docs/tool-profiles.md
@@ -17,10 +17,12 @@ a profile is a managed policy, a grant attaches it to a principal (the session),
 > `exec*`, but that takes away a convenient entry point, not a privilege
 > ([#45](https://github.com/openabdev/instance-mcp/issues/45)). When you grant full control, lend
 > a **dedicated computer** (a Linux hands node, a throwaway machine or VM), not the one you work
-> on. Every served tool is classified `observe` (reads only), `act` (changes state) or `shell`
-> (reaches the desktop user's shell); the boundary tests fail on any **unclassified** tool, on any
+> on. Every **local** tool is classified `observe` (reads only), `act` (changes state) or `shell`
+> (reaches the desktop user's shell); the boundary tests fail on any **unclassified** local tool, on any
 > profile that claims to be narrower than a shell while allowing a `shell` tool, and on `observe`
 > holding anything but `observe` tools (macOS `ProfileBoundaryTests`, Linux `profile_tests`).
+> Upstream `browser_*` tools are governed by their own allowlist (`desktop` gets 15, `observe`
+> none) and are not in that classification yet; they must be before a `browser` tier ships.
 
 ## 1. Available profiles
 
@@ -106,8 +108,10 @@ change anything.
 **`observe` protects the computer, not the agent.** Every screenshot goes into the agent's
 context. If the screen shows a hostile page or message, its text can act as a prompt injection
 against the agent — which cannot act on this computer under `observe`, but still has its own shell
-in the pod, its model API, and whatever credentials the session holds (for example git). Treat what
-is on screen as untrusted input to the agent.
+in the pod, whatever credentials the session holds (for example git), and **outbound internet
+access by default**: it can send whatever it reads to any host. Treat what is on screen as
+untrusted input to the agent; deployments that need to contain this should give the pod an egress
+allowlist (model API, git remotes, package registries).
 
 ## 3. What each profile loses against the previous one