Report vulnerabilities privately to the maintainers (do not open public issues).
Untrusted: the website and its prompts, the relay/network, agent and tool output, deep links, custom adapters.
Implemented controls:
- Outbound-only desktop. No listening sockets.
http://only for localhost relays. - Service authentication.
Bearer <serviceId>.<secret>; secrets stored as SHA-256, compared in constant time. - Desktop authentication. Per-desktop random token stored in an owner-only (0600) file in the app data folder (same model as
~/.codex/~/.claude); the OS keychain is not used, so unsigned rebuilds don't trigger authorization prompts. - Pairing. 256-bit single-use tickets (or 50-bit codes, rate limited to 5 failures/min/desktop), 30–600 s TTL, bound to the first desktop that redeems them. Service identity on the consent screen comes from the relay registry, never from the link. Unverified services get a red warning.
- Consent. Explicit approval per connection with folder, access level, duration. Per-task approval on by default. Sensitive tool actions prompt locally; the website can only observe. "Allow for this task" is only offered for read/edit/search/fetch kinds.
- Tenant isolation. A service can only see its own pairings, connections and tasks; tasks must name the same external user the desktop consented to (404 otherwise).
- Workspace confinement for ACP fs methods: absolute paths only, canonicalization, symlink and hard-link escape checks, credential folder denylist (
.ssh,.aws,.gnupg,.claude,.codex, …), size caps. Home folder and/cannot be chosen as a workspace. - Output firewall (website tasks). Real adapters read files with their own tools (not ACP
fs/*), and Codexread-onlycan read the whole disk without asking. The bridge therefore inspects every tool call (title, raw input, locations) and permission request: any path outside the project folder in home/system/temp areas auto-denies the permission or stops the task, and nothing more is forwarded. Outgoing text is scanned (rolling window) for private keys and common token formats. Verified with real codex-acp 2.5.0 and claude-agent-acp 0.100.0 (tests/e2e/tests/real_agents.rs). The agent's own skill folders (~/.agents/skills,~/.codex/skills,~/.claude/skills, skills inside~/.claude/plugins) may be read by non-mutating tools, because agents load the skills the user installed; credential files are outside those folders and stay blocked. Heuristic: an obfuscated shell command can evade path detection; the agent's own restricted mode and local permission prompts remain the primary control. - Website instructions and tools. Instructions are appended to the agent's system prompt (ACP
systemPrompt) and are shown to the user on the task; they cannot change desktop policy. Website tools (MCP over HTTP) require the user-consentedagent.toolsscope, at most 8 servers, https, and a host equal to (or a subdomain of) the service's verified origin. This is checked on the desktop and prevents pointing the agent at localhost/LAN services. Header values go only to the local agent. - Agent sandbox. Remote sessions are switched to a restricted adapter mode (Codex
read-only/workspace-write, Claudedefault/acceptEdits) or refused; permissive modes (auto,agent,bypassPermissions,*full-access*) are never selected, and a task is stopped if the agent switches into one. Terminal capability is not advertised. - Process control. Agents run with a cleared, allowlisted environment, in their own process group; cancel →
session/cancel→ kill tree after 10 s. - Kill switches. Pause bridge, pause/revoke a website, revoke all — effective immediately on both desktop and relay.
- No credential handling. Provider OAuth caches are never read. Sign-in is always the official CLI flow (
codex login,claude auth login); the app has no API-key handling. - Redaction. Workspace/home paths and token-like strings are scrubbed from titles sent to websites and from diagnostics. Account identity notifications from adapters are never forwarded.
- Limits. 1 MiB frames, 64 KiB prompts, bounded queues/buffers, rate limits per service, max task duration, approval timeout, desktop-offline grace.
- Audit log of pairing, grants, tasks, permission outcomes, fs denials.
Known gaps are tracked in PROGRESS.md (“Not done yet”).