Vulnerabilidades do Openclaw Openclaw

OpenClaw (npm package `openclaw`) versions >= 2026.7.2 and < 2026.9.2 contain a race condition in the Discord realtime voice transcript path. Concurrent control-classified voice transcripts could consume speaker context belonging to another participant after an asynchronous control check, causing a transcript to inherit another speaker's owner status. In Discord agent-proxy voice sessions using the affected realtime control path, an utterance from a non-owner participant could reach the downstream agent boundary marked as owner, so owner-sensitive behavior is applied to the wrong speaker. Exploitation depends on concurrent transcript timing and on the tools and commands available to the affected agent. The issue is fixed in 2026.9.2; as a workaround, disable Discord realtime voice for agents that distinguish owner and non-owner senders.

OpenClaw (npm package `openclaw`) before 2026.8.1 incorrectly enforces sender tool policies during session-memory filename generation. In affected versions, filename generation created an embedded helper that retained tools which the originating sender's policy had removed. When session-memory filename generation was enabled for an agent reachable by lower-trust senders, model-mediated instructions could cause the helper to invoke tools outside that sender's effective policy; the demonstrated impact was the creation of persistent scheduled work. Exploitability depends on the model acting on the injected instruction and on which tools the helper exposes. The issue is fixed in 2026.8.1; as a workaround, disable session-memory filename generation for agents reachable by lower-trust senders.

OpenClaw (npm package 'openclaw') before 2026.8.1 does not apply the originating sender's global or per-agent toolsBySender policy when handling outbound attachments. A sender that has been explicitly denied filesystem read tools can still cause a known local file to be read and returned via a final-response media directive or a message attachment, disclosing local file contents to an admitted requester whose agent turn did not include the read tool. Exploitation requires knowledge or derivation of a useful host path and a delivery flow that accepts local attachments. The issue is fixed in version 2026.8.1.

OpenClaw versions before 2026.8.1 fail to validate all source fields in structured message attachments, allowing attackers to hide unvalidated host paths behind allowed attachment sources. Attackers can exploit this by providing multiple source fields to bypass sandbox path validation and cause Telegram delivery to read and send known host files that would otherwise be rejected.

OpenClaw versions before 2026.8.1 contain a path traversal vulnerability in the tools.fs.workspaceOnly feature where Unicode filename fallback can normalize validated parent directory components. Admitted requesters can exploit canonically equivalent sibling directories to read files outside the configured workspace boundary.

OpenClaw versions before 2026.8.1 contain an authorization scope widening vulnerability in file-transfer allow-always approvals that allows attackers to reuse standing grants for unreviewed paths. Attackers can exploit glob metacharacter interpretation and node display name reuse to access sibling paths or different nodes beyond the operator's original approval scope.

OpenClaw (npm package 'openclaw') before 2026.8.1 could send third-party provider credentials to the wrong endpoint. In affected versions, when a third-party provider uses an OpenAI-compatible API and the resolved model metadata lacks a concrete base URL, a pinned session that continues after a model configuration hot reload retains that provider's credential while the OpenAI SDK selects its own default endpoint. A resulting request could disclose the configured third-party provider credential to an unrelated provider endpoint and fail with a misleading authentication error. Operators who observed this condition should rotate the affected credential. The issue is fixed in 2026.8.1.

OpenClaw before 2026.8.2 contains a denial of service vulnerability in the Browser extension relay that allows unauthenticated network sources to exhaust pending-authentication capacity. Attackers can hold every pending slot by maintaining silent WebSocket upgrades, preventing paired extensions from completing Browser Relay Authentication v2.

OpenClaw versions before 2026.5.18 contain an authorization bypass vulnerability in skill command dispatch that allows lower-trust callers to execute or persist actions beyond their intended authorization. Attackers can bypass tool policy restrictions through configured input paths to perform unauthorized actions when the affected feature is enabled and reachable.

OpenClaw 2026.5.12 before 2026.5.26 contain an incorrect authorization vulnerability in the ClickClack allowFrom feature. When the affected feature is enabled and reachable, a lower-trust caller or configured input path could execute or persist actions beyond the caller's intended authorization, including running non-allowlisted commands.

OpenClaw 2026.2.25 before 2026.5.26 allow a lower-trust caller or configured input path to bypass non-browser rate limits on WebSocket authentication attempts. When the affected feature is enabled and reachable by lower-trust input, this can consume gateway resources and reduce service availability.