OpenClaw vulnerabilities

OpenClaw (npm package 'openclaw') before 2026.7.1 is vulnerable to a time-of-check time-of-use race condition in OpenShell local mirror filesystem mutation operations. The remove, mkdir, and rename operations could act on a different filesystem target after OpenClaw completed its sandbox path-safety check, if the path is changed concurrently. An attacker able to win the race can cause a sandboxed operation to delete, create, or rename a host path outside the intended mirror root with the permissions of the OpenClaw process user. This does not require an operator to have granted host filesystem access outside the sandbox. Fixed in 2026.7.1.

OpenClaw versions before 2026.7.1 fail to properly authorize non-owner users executing MCP configuration changes through /mcp set and /mcp unset commands. Attackers can persist arbitrary stdio MCP commands that execute with OpenClaw process privileges when configuration loads, compromising host confidentiality, integrity, and availability.

OpenClaw versions before 2026.7.1 contain an authorization bypass vulnerability in the diagnostics export command that allows non-owner channel senders to access owner-only host diagnostic bundles. Attackers can request and receive diagnostic details about the host, configuration, runtime, and connected services intended only for owners.

OpenClaw versions before 2026.7.1 contain an authorization bypass vulnerability in the /export-trajectory endpoint that allows non-owner senders to request and receive owner-only trajectory bundles. Attackers can access prompts, model messages, tool schemas, runtime events, and local path metadata from affected sessions by exploiting insufficient authorization checks.

OpenClaw (npm package `openclaw`) before 2026.7.1 does not enforce the documented owner-only requirement for persistent `/activation` policy changes in group channels. An authorized non-owner channel sender can change whether the agent requires mention-based activation, causing the agent to respond more broadly in the group (exposing its responses to additional group traffic) or suppressing expected activation behavior until an owner restores the intended setting.

OpenClaw is an agent gateway distributed via npm. In versions >= 2026.4.10 and < 2026.7.1, persistent memory dreaming mutations omit owner permission checks. An authorized but non-owner external-channel sender can issue the persistent '/dreaming on' and '/dreaming off' commands to enable or disable the Gateway's Memory Core dreaming behavior, disabling background memory processing or re-enabling durable memory promotion where the owner expected it to remain disabled; the practical confidentiality, integrity, and availability impact depends on stored conversation material and subsequent memory use.

OpenClaw is an npm-distributed agent gateway. In versions before 2026.7.1, the global Active Memory toggle mutations could omit owner checks. An authorized non-owner external-channel sender could therefore persistently enable or disable Active Memory for the Gateway, disabling memory recall for future sessions or re-enabling global recall where the owner expected it to remain disabled. The issue is limited to persistent global on/off mutations; session-level controls and read-only status remain governed by existing command policy.

OpenClaw before 2026.7.1 contains an authorization bypass vulnerability in the /voice set command that allows non-owner external-channel senders to persist Gateway voice configuration. Attackers with command access can change the voice used by Talk responses for the configured provider, affecting configuration integrity without exposing credentials or granting additional host capabilities.

OpenClaw versions before 2026.7.1 contain a sandbox bypass vulnerability in the browser tool that allows sandboxed sessions to access paired node browser actions despite allowHostControl=false configuration. Attackers with control over sandboxed agent input can select a paired node and perform host browser operations, inspecting or manipulating the connected browser profile and its authenticated state.

OpenClaw before 2026.7.1 does not enforce the administrator scope requirement on browser control when reached through the node.invoke method, although direct browser.request access requires administrator scope. In Gateway deployments that honor caller identity, a write-scoped caller can inspect pages, navigate tabs, or interact with browser-visible applications without the configured admin requirement. The issue is fixed in 2026.7.1.

OpenClaw before 2026.7.1 fails to properly validate owner authorization in the Codex computer-use installation command. Non-owner channel senders can install arbitrary plugins and execute MCP processes with OpenClaw user privileges, affecting host confidentiality, integrity, and availability. The issue is fixed in 2026.7.1.

OpenClaw before 2026.7.1 fails to enforce the owner-only authorization requirement for Claude Code permission prompts delivered through the MCP channel bridge. An authorized non-owner channel sender with channel command access can approve or deny a pending permission request intended for the owner, causing the requested action to proceed without owner consent. The issue is fixed in 2026.7.1.

OpenClaw is an npm-distributed agent runtime. In versions >= 2026.2.26 and < 2026.7.1, PowerShell command analysis on Windows hosts running in exec allowlist mode could approve an exact executable resolved from PATH but subsequently execute a same-named executable located in the workspace directory. If lower-trust content can place an executable with an approved basename into an agent-writable workspace and steer an approved PowerShell command that uses a bare executable name, OpenClaw may run the workspace file instead of the allowlisted path, executing arbitrary code with the privileges of the Gateway or node-host user.

OpenClaw Discord versions before 2026.7.1 contain an authorization bypass vulnerability in guild metadata read actions that allows lower-trust senders to retrieve information excluded by channel allowlists. Attackers can bypass the configured Discord read-target policy to access guild metadata from servers or channels outside the operator's allowlist.

OpenClaw channel plugins (@openclaw/msteams, @openclaw/feishu, @openclaw/matrix, and @openclaw/googlechat) before 2026.8.1 do not enforce the configured channel read allowlist for caller-supplied explicit read targets in message, reaction, pin, member, and related metadata read actions. A lower-trust sender or a steered agent with access to a channel read action can therefore retrieve content or metadata from channels or rooms excluded by the operator's read policy; the practical impact depends on the permissions held by the connected bot account.

OpenClaw for iOS before 2026.8.11 stores Gateway credentials as cleartext JSON in App Group UserDefaults instead of the device Keychain. Attackers with access to unencrypted device backups or extracted App Group containers can recover valid Gateway tokens and passwords to authenticate with operator authority.

OpenClaw before 2026.7.1 improperly handles case sensitivity in the model-facing cron tool: a mixed-case payload kind can pass the agent-facing shell-execution guard and later normalize into a command job. An actor able to steer a tool-enabled agent can create a persistent cron job that executes attacker-selected commands with the privileges of the OpenClaw process user. The issue is fixed in 2026.7.1.

OpenClaw before 2026.7.1 incorrectly trusts requester provenance in message.action. In identity-bearing Gateway deployments, a write-scoped caller can supply another sender's identifier to the channel authorization checks and invoke a channel action under that spoofed requester identity, reaching operations the channel adapter would have denied to the real caller. The issue is fixed in 2026.7.1.

OpenClaw before 2026.7.1 fails to restrict owner-only infrastructure tools exposed through the chat.send endpoint. In Gateway deployments using authentication modes that honor caller identity, a write-scoped non-owner caller can start a chat turn whose tool inventory includes the gateway and cron tools, causing the agent to invoke owner-only configuration or scheduling operations, including persistent state changes. The issue is fixed in 2026.7.1.

OpenClaw versions before 2026.8.1 fail to validate video asset URLs returned by providers, allowing server-side requests to private destinations. A malicious or compromised provider can return private or loopback URLs to cause the CLI to make requests to internal services accessible from the OpenClaw host.

OpenClaw versions before 2026.8.1 contain a server-side request forgery vulnerability in browser wait predicates that allows attackers to bypass SSRF protections by reaching blocked destinations. Attackers can use the wait --fn function against an existing browser session to request loopback or private destinations without navigation checks applied to other browser actions.

OpenClaw (npm package 'openclaw') before 2026.8.1 contains a server-side request forgery vulnerability in its trusted-host DNS checks. For fetches that use the trusted-host DNS recheck, a trusted hostname that resolves to an unspecified address (0.0.0.0 or ::) bypasses the SSRF destination validation. An attacker who can influence DNS for an allowed hostname can therefore cause a guarded fetch to reach a service bound only to loopback and disclose its response; the practical impact depends on the reachable service and the data it returns. Fixed in 2026.8.1.

OpenClaw versions before 2026.8.1 contain a sandbox policy bypass vulnerability in the MCP loopback component that allows sandboxed coding-agent sessions to invoke tools explicitly denied by sandbox.tools.deny policy. Attackers can list and invoke denied tools to access data or perform actions the operator intended to exclude from the sandbox.

OpenClaw versions >= 2026.3.25 and < 2026.8.1 apply invalid-token rate limiting for Synology Chat webhooks before authentication and key the limit on the raw proxy socket address. In deployments where OpenClaw sits behind a trusted reverse proxy or tunnel and multiple external clients share a single socket address, an unauthenticated sender can exhaust the shared invalid-token budget, causing subsequent legitimate Synology Chat webhook callbacks to be rejected until the rate-limit window expires. The attacker cannot obtain a valid token or read message data; the impact is temporary loss of channel availability. Fixed in 2026.8.1.

OpenClaw (npm package 'openclaw') versions >= 2026.6.6 and < 2026.8.1 apply the SMS webhook invalid-request rate limit before Twilio signature verification and identify clients only by the raw proxy socket address. In deployments where the SMS/Twilio webhook sits behind a trusted reverse proxy or tunnel so that multiple external clients share one socket address, an unauthenticated remote sender can exhaust the shared pre-authentication rate-limit budget with invalid requests, causing subsequent correctly signed Twilio callbacks to receive HTTP 429 responses. The result is temporary inbound SMS loss while the shared window remains exhausted; the attacker cannot forge a callback or access message data. Fixed in 2026.8.1.

OpenClaw versions >= 2026.3.28 and < 2026.8.1 allow an untrusted workspace .env file to set the CLOUDSDK_PYTHON_ARGS environment variable. When an operator starts OpenClaw in attacker-controlled workspace content and then runs the Gmail setup flow, that value is inherited when gcloud is launched, and the gcloud launcher passes it as arguments to the trusted Python interpreter. The issue is fixed in 2026.8.1.

OpenClaw is an npm-distributed application. In versions >= 2026.4.25 and < 2026.8.1, the workspace environment-variable filter did not block variables ending in `_ENDPOINT`, so an untrusted workspace `.env` file could set AZURE_SPEECH_ENDPOINT. Azure Speech preferred that value over the configured region, so when a synthesis or voice-list request was made, the attacker-selected endpoint received the operator's Azure Speech key in the request header, allowing the key to be reused against the operator's Azure Speech resource. Exploitation requires an operator to start OpenClaw in attacker-controlled workspace content with Azure Speech configured with a key and region and no trusted endpoint override set. The issue is fixed in 2026.8.1.

OpenClaw before 2026.8.1 fails to properly restrict access to operator command cron jobs, allowing model-visible agent callers to read and execute ownerless command jobs. Attackers can inspect stored environment variables and force-run disabled or unscheduled command jobs to access secrets and execute operator-authored commands. The issue is fixed in 2026.8.1.

OpenClaw versions >= 2026.4.5 and < 2026.8.1 validate a single DNS resolution result for a configured remote Chrome DevTools Protocol (CDP) hostname, but the WebSocket and Playwright transports perform a later, independent DNS resolution, discarding the DNS pinning enforced at validation time. An attacker controlling an approved CDP hostname or its DNS answers can exploit this check-then-use gap via DNS rebinding to make the Gateway connect to a loopback, private, link-local, cloud metadata, or other SSRF-policy-denied address.

OpenClaw LINE versions before 2026.8.1 contain an access control vulnerability where group allowlist mode silently inherits DM allowFrom values when groupAllowFrom is not explicitly configured. Attackers with group participation can trigger the agent despite configured group allowlist restrictions when DM access is broader than intended group access.

OpenClaw versions before 2026.8.1 fail to neutralize spreadsheet formula characters in participant display names within attendance CSV exports. Attackers can inject formula-like cells that execute with spreadsheet user permissions when the export is opened in formula-enabled applications.

OpenClaw (npm package `openclaw`) before 2026.8.1 does not neutralize leading characters that spreadsheet applications interpret as formulas when the Control UI exports session data to CSV. Although session labels were quoted as CSV text, a lower-trust participant who can influence a session label or the first user message can place a formula-like cell in the Usage export; if an operator opens that export in a spreadsheet application with formula evaluation enabled, the cell may be evaluated with the permissions of the spreadsheet user. OpenClaw itself does not evaluate the formula, and practical impact depends on the spreadsheet application's security settings. The issue is fixed in 2026.8.1.

OpenClaw versions before 2026.8.1 contain an authorization bypass vulnerability in the sessions.create endpoint. It allows operator.write callers to modify session configurations reserved for the operator.admin scope.

OpenClaw (npm package 'openclaw') versions >= 2026.3.22 and < 2026.8.1 contain an approval-bypass flaw in the exec approval policy: the policy could trust a command-running wrapper without inspecting the command carried in its arguments. After an operator allowlisted or permanently approved a benign wrapper invocation, a later agent turn could substitute an arbitrary inner command and execute it with the OpenClaw process's host privileges without a further approval prompt.

OpenClaw versions before 2026.8.1 contain an authorization bypass vulnerability where Allow Always approvals for exact commands persist as path-only grants on macOS and Linux. Attackers can reuse the same executable with different arguments to execute commands without triggering new approval prompts, potentially accessing files or internal services.

OpenClaw versions before 2026.8.1 contain a command parser vulnerability where escaped newlines confuse exec allowlist parsing, allowing hidden commands to execute. Attackers can craft input with escaped newlines to bypass allowlist validation and execute additional commands without expected authorization prompts.

OpenClaw versions before 2026.8.1 contain a resource exhaustion vulnerability in the Gateway listener that allows unauthenticated clients to retain response sockets by sending WebSocket upgrade requests without matching connection semantics. Attackers can repeatedly send malformed upgrade requests to exhaust listener resources and cause denial of service without consuming the WebSocket pre-auth connection budget.

OpenClaw versions before 2026.8.1 contain an authorization bypass vulnerability in skill tool dispatch that fails to carry the sender's owner status. Non-owner senders authorized to invoke skill commands can access owner-only tools and server credentials reserved for owners.

OpenClaw (npm package openclaw) versions >= 2026.5.2 and < 2026.8.1 contain an incorrect authorization vulnerability in WhatsApp group handling. A group sender admitted for ordinary messages but denied by commands.allowFrom or owner command authorization can issue the /new <model> command to reset the shared group session and persist a provider and model override.

OpenClaw is an npm-distributed gateway application. In versions >= 2026.7.1 and < 2026.8.1, Synology Chat attachment delivery could lose DNS pinning: the Gateway validated a single DNS result for a supplied file URL but then passed the original hostname to the Synology NAS, where it could resolve to a different destination. When attachment delivery accepted a remotely influenced hostname, an attacker could use DNS rebinding to make the NAS fetch a private or otherwise policy-denied resource and return its contents to the addressed conversation (server-side request forgery).

OpenClaw (npm package 'openclaw') versions >= 2026.5.12 and < 2026.8.1 do not immediately invalidate Canvas HTTP authorization when a paired node is revoked. Node revocation invalidates the WebSocket client, but Canvas HTTP authorization continues to accept and renew the previously granted capability until WebSocket close cleanup completes.

OpenClaw versions >= 2026.6.9 and < 2026.8.1 do not declare the native chatId parameter as a delivery target in the Feishu unpin feature, so unpin requests can bypass the shared same-provider cross-context target check. When tools.message.crossContext.allowWithinProvider is disabled, an admitted (authenticated) sender can remove a pin from another Feishu group that the sender and the configured account are otherwise permitted to access.

OpenClaw (npm package 'openclaw') before 2026.8.1 does not correctly enforce per-chat tool policies for Codex app-server runtime tools. A conversation-level tools.allow rule filtered OpenClaw tools but did not restrict the shell, process, file, and patch tools owned by the Codex runtime. When a lower-trust conversation was assigned to a Codex runtime and restricted with a per-chat tool allowlist, a participant able to trigger that agent could still reach native command and file tools, bypassing the configured allowlist.

OpenClaw (npm package 'openclaw') before 2026.8.1 contains an access-control bypass in the Microsoft Teams integration. When groupPolicy is set to allowlist, a missing or unsupported configured access group produces a denied group-resolution result that is not rejected by the final message-admission check.

OpenClaw versions before 2026.8.1 contain a path traversal vulnerability in QQBot voice attachment handling where filenames are decoded twice, allowing encoded traversal segments to reappear after sanitization.

OpenClaw (npm package 'openclaw') versions >= 2026.3.28 and < 2026.8.1 contain a credential exposure issue in memory embedding failover. When remote embedding fallback is configured and the primary embedding provider fails, the request can be sent to a different fallback provider while still reusing the primary provider's configured API key.

OpenClaw is a coding agent distributed as the npm package `openclaw`. In affected versions (2026.7.1 through 2026.7.2), alternate but valid `file:` URL spellings supplied over the Agent Client Protocol (ACP) were treated as relative paths and were incorrectly classified as reads scoped to the session working directory.

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.

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.

Before version 2026.8.1, the npm package openclaw could include deterministic hashes computed over the original, unredacted configuration in redacted configuration responses. When the Gateway password had low entropy and the remaining configuration values were reconstructable, these hashes acted as offline password verifiers: a caller able to obtain the redacted configuration (for example via config.get) could test password candidates offline without going through the rate-limited Gateway authentication path. Recovering the password could grant the documented shared-secret operator authority. Secret references were not affected in the same way. The issue is fixed in 2026.8.1.