CVE-2026-35620

Description

> Fixed in OpenClaw 2026.3.24, the current shipping release. **Title** Non-owner command-authorized sender can change the owner-only `/send` session delivery policy **CWE** CWE-285 Improper Authorization **CVSS v3.1** CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:L/A:L Base score: **5.4 (Medium)** **Severity Assessment** Medium. This is a real owner-only authorization bypass, but the demonstrated impact is limited to persistent mutation of the current session’s delivery policy rather than direct code execution, sandbox escape, or cross-host compromise. **Impact** A non-owner sender who is allowed to run commands can invoke `/send on|off|inherit` and persistently change the current session’s `sendPolicy`, even though OpenClaw documents `/send` as owner-only. That lets a lower-trust participant: - disable reply delivery for the current session (`/send off`), suppressing future replies in that chat; - re-enable reply delivery (`/send on`) after the owner intentionally disabled it; - remove the session override (`/send inherit`). **Affected Component** Verified against the latest published GitHub release tag `v2026.3.23` (`ccfeecb6887cd97937e33a71877ad512741e82b2`), published `2026-03-23T23:15:50Z`. Exact vulnerable path on the shipped tag: - `src/auto-reply/reply/commands-session.ts:212-239` - `handleSendPolicyCommand(...)` checks only `params.command.isAuthorizedSender`. - when true, it mutates `params.sessionEntry.sendPolicy` and persists the session entry. Authorization behavior that makes this reachable: - `src/auto-reply/command-auth.ts:401-407` - `senderIsOwner` is computed separately from general command authorization. - `src/auto-reply/command-auth.ts:420-429` - command authorization can succeed even when `senderIsOwner === false`. - `src/auto-reply/command-auth.owner-default.test.ts:10-47` - existing coverage confirms a sender can be command-authorized while not treated as owner. Documented owner-only contract: - `docs/tools/slash-commands.md:112` - `/send on|off|inherit` is documented as owner-only. - `docs/concepts/session-tool.md:156` - `sendPolicy` is documented as settable via `sessions.patch` or owner-only `/send on|off|inherit`. Related privilege model: - `src/gateway/method-scopes.ts:131-133` - `sessions.patch` is admin-scoped, which reinforces that session-delivery-policy mutation is treated as privileged state. Version history: - The vulnerable handler exists in release history going back at least to commit `ea018a68ccb92dbc735bc1df9880d5c95c63ca35` (`refactor(auto-reply): split reply pipeline`). - Earliest released affected tag found: `v2026.1.14-1` - Latest released affected tag verified: `v2026.3.23` **Technical Reproduction** 1. Check out the shipped release tag `v2026.3.23`. 2. Configure a channel where: - a non-owner sender is allowed to run commands, for example through `commands.allowFrom`; - the owner identity is distinct, for example via `commands.ownerAllowFrom`. 3. Start or reuse a session with a live `sessionEntry` and `sessionStore`. 4. Send `/send off` as the non-owner but command-authorized sender. 5. Confirm the resolved command context has: - `isAuthorizedSender === true` - `senderIsOwner === false` 6. Observe that the handler still accepts the command, mutates `sessionEntry.sendPolicy`, and persists the session entry. **Demonstrated Impact** The vulnerable handler performs a real persistent session-state change: - `src/auto-reply/reply/commands-session.ts:232-238` - `/send inherit` deletes `sessionEntry.sendPolicy` - other modes assign `sessionEntry.sendPolicy = sendPolicyCommand.mode` - the handler then calls `persistSessionEntry(params)` The mutation is not gated by owner status, only by general command authorization. That changes subsequent delivery behavior for the current session, which matches the documented meaning of `sendPolicy`. **Environment** - Product: OpenClaw - Verified shipped tag: `v2026.3.23` - Shipped tag commit: `ccfeecb6887cd97937e33a71877ad512741e82b2` - Published GitHub release time: `2026-03-23T23:15:50Z` - Verification date: `2026-03-24` **Duplicate Check** Upon inspection there is no preexisting GHSA for `/send`. This is distinct from: - `GHSA-r7vr-gr74-94p8` - that advisory covered owner-only authorization bypasses for `/config` and `/debug`, not `/send`. This is the same authorization class, but a different privileged command surface that still lacks the owner check. **In Scope Check** This report is in scope under `SECURITY.md` because: - it does **not** rely on adversarial operators sharing one gateway host or config; - it does **not** rely on trusted local state tampering; - `SECURITY.md:151-152` explicitly says non-owner sender status matters for owner-only tools and commands; - `/send` is explicitly documented as owner-only, so this is a direct owner-only authorization bypass, not a complaint about normal shared-agent steering. This is therefore a concrete authorization flaw against a documented product boundary. **Remediation Advice** 1. Change `/send` to require owner status, not just command authorization. 2. Reuse the same owner-only rejection pattern already used by privileged command surfaces such as `/config`, `/debug`, and owner-only `/plugins` writes. 3. Add regression coverage for the exact case where: - a non-owner sender is command-authorized; - `/send` must still be rejected unless `senderIsOwner === true`. 4. Verify that the owner can still use `/send on|off|inherit` normally.