OpenClaw CLI versions from 2026.1.5 up to and including 2026.2.21-2 contain a vulnerability in the shell environment fallback mechanism. The issue involves trusting environment variables such as SHELL, HOME, and ZDOTDIR during startup, which could allow execution of attacker-influenced login shell startup paths before environment keys are loaded. This can lead to local code execution if an attacker can influence environment or configuration input. The vulnerability is not exploitable remotely and depends on crossing local trusted operator boundaries. A fix is available in the mainline and planned for release in version 2026.2.22, which hardens environment variable handling and sanitizes the shell fallback execution environment.
The OpenClaw CLI has a vulnerability in its webhook transform modules where path validation only uses lexical checks. This allows a symlink within the allowed transform directory to resolve outside the intended directory, enabling dynamic import of unintended JavaScript code. The issue affects versions up to and including 2026.2.21-2 and is planned to be fixed in version 2026.2.22. Exploitation requires that hook transforms are enabled and reachable, and that an attacker can influence the transform path resolution to reference a symlink escape to attacker-controlled code. The vulnerability could lead to execution of unauthorized code with gateway-process privileges.
OpenClaw CLI versions up to 2026.3.2 have a vulnerability where non-POST HTTP requests to the /hooks/* endpoint are counted toward the authentication failure limit, allowing an unauthenticated attacker to trigger a temporary lockout of webhook delivery. This can cause temporary denial of service for legitimate webhook-triggered automation. The issue is fixed in version 2026.3.7, which correctly rejects unsupported HTTP methods with a 405 status without incrementing the auth failure count.
OpenClaw CLI versions up to 2026.2.22-2 contain a vulnerability where exec.approval requests for host=node were not bound to a specific nodeId. This allowed an operator's approval for a system.run request on one node to be replayed on a different node within the same gateway fleet. The issue is fixed in version 2026.2.23 by requiring and persisting nodeId for approval requests and rejecting execution if the node binding does not match. Users should upgrade to 2026.2.23 or later to remediate this issue.
OpenClaw CLI versions up to 2026.2.22-2 have a vulnerability where the exec allowlist and safe-bins policy evaluation can diverge from actual runtime execution due to how GNU env -S wrapper commands are interpreted. This mismatch allows an attacker who can influence command text to bypass allowlist restrictions and execute unintended commands. The issue is fixed in version 2026.2.23 by enforcing stricter execution planning and policy checks.
OpenClaw CLI versions up to 2026.2.22-2 contain a vulnerability in a non-default configuration where the autoAllowSkills setting is enabled. This flaw allows a skill-name collision to bypass the 'ask=on-miss' execution approval prompt, potentially permitting execution without operator confirmation. The issue arises because the allowlist evaluator matches skill names by basename, allowing path-scoped executables to bypass prompts. The vulnerability is fixed in version 2026.2.23 by requiring pathless invocation tokens and trusted resolved executable paths for auto-allow matching.
A vulnerability in the openclaw npm package (openclaw-cli) affects Microsoft Teams media fetch paths by bypassing the shared SSRF guard model. This inconsistency in host and DNS enforcement across redirect and fetch hops could lead to SSRF boundary policy drift. The issue is fixed by unifying fetch paths through a shared SSRF-guarded mechanism with centralized hostname allowlist policies. Versions up to 2026.2.25 are affected, with a patch planned in version 2026.2.26.
OpenClaw CLI versions up to 2026.2.26 contain a time-of-check/time-of-use (TOCTOU) vulnerability in sandbox media handling. This flaw allows a symlink attack to cause media reads outside the intended sandbox root directory, potentially exposing host files. The issue arises because media paths are validated at one point and read later through a different path, allowing a race condition. The vulnerability is fixed in version 2026.3.1 by consolidating media read paths to be root-scoped and boundary-safe, eliminating the TOCTOU gap.
OpenClaw CLI versions prior to 2026.4.2 contain a vulnerability in the QQ Bot structured media payload handling that allows reading arbitrary local files on the host. This occurs because crafted payloads can escape the intended media root directories, leading to unauthorized file reads. The issue is fixed in version 2026.4.2 by restricting local path access in structured payloads.
OpenClaw CLI versions from 2026.1.20 through 2026.2.17 contain a Server-Side Request Forgery (SSRF) guard bypass vulnerability. The SSRF protection failed to detect ISATAP embedded IPv4 addresses within IPv6 literals, allowing crafted URLs to bypass private IP filtering. This weakness permits SSRF attempts targeting internal or private network addresses via URL fetching paths. The issue is rated medium severity and has been fixed in version 2026.2.19 by adding ISATAP IPv4 detection and centralizing hostname/IP blocking.
OpenClaw CLI versions up to 2026.2.26 have an authorization bypass vulnerability allowing unauthorized senders to trigger stop-like abort commands and access the /models command output without proper sender authorization. This could disrupt active sessions and expose model and authorization metadata that should be restricted. The issue is fixed in version 2026.3.1 by enforcing sender authorization on these command paths.
OpenClaw CLI versions up to 2026.2.21-2 have a vulnerability in the tools.exec.safeBins feature where a generic fallback safe-bin profile allowed interpreter-style binaries to execute inline payloads in allowlist mode. This issue requires explicit operator configuration to add such binaries, limiting impact to misconfigured deployments. A fix is planned in version 2026.2.22 that removes the generic fallback and requires explicit safe-bin profiles.
OpenClaw Gateway's authenticated HTTP endpoint /tools/invoke allowed invocation of high-risk session orchestration tools by default, enabling escalation from single-tool invocation to session spawning and control. Additionally, ACP clients could auto-approve risky permission requests without sufficient user interaction, increasing risk of silent execution or mutation. The vulnerability affects OpenClaw versions prior to 2026.2.14 and has been fixed in 2026.2.14 with hardened default behaviors and stricter permission handling.
In sandboxed runs of the openclaw-cli package, native prompt image auto-load did not respect the optional hardening setting tools.fs.workspaceOnly=true. This allowed prompt text to reference and load image files from mounted paths outside the intended workspace, potentially exposing sensitive image data to vision-capable model inputs. This issue affects versions before 2026.2.24 and requires specific conditions including sandbox mode, the workspaceOnly setting enabled, an out-of-workspace mount, and active vision-capable model paths. The vulnerability has been fixed in version 2026.2.24.
OpenClaw CLI versions up to 2026.2.21-2 contain a symlink traversal vulnerability in the gateway session metadata handling. A crafted local avatar path can follow symlinks outside the agent workspace, allowing an attacker to read arbitrary local files and return their contents as base64 data URLs in gateway responses. This leads to local file disclosure within the gateway process context. A patch is planned in version 2026.2.22 to fix this issue.
OpenClaw CLI versions up to 2026.3.2 have a vulnerability in the system.run allowlist persistence logic where shell comment semantics were not properly honored. This allowed non-executed command tails following unquoted '#' characters to be stored as trusted commands, potentially leading to unintended allowlist persistence. The issue was fixed in version 2026.3.7 by improving shell tokenization to correctly handle unquoted shell comments.
The Openclaw CLI's Google Chat integration allows authorization based on mutable email principals in its allowlist, weakening identity binding. This can lead to spoofing access if an attacker has Google Workspace or IdP administrative control to reassign email addresses. The issue is low severity because requests are authenticated via Google Chat tokens, and exploitation requires high privilege access. The vulnerability affects openclaw versions from 2026.1.29 up to but not including 2026.2.14. A patch is available that deprecates treating 'users/<email>' as an email allowlist entry and recommends using immutable user IDs instead.
OpenClaw CLI had a security hardening update addressing canvas route access in mixed-trust or network-visible deployments. Previously, fallback authorization behavior could allow broader access than intended. The update requires explicit token or session-capability authorization and removes shared-IP fallback paths, tightening access controls. The risk is mainly for deployments exposing canvas routes beyond loopback or trusted hosts. Versions up to 2026.2.19-2 are vulnerable, with a patch targeted for 2026.2.21.
OpenClaw CLI versions up to 2026.2.21-2 have an authorization issue where the elevated allowFrom configuration accepted broader identity signals than intended, potentially allowing elevated sender approval based on less strict identity matching. The fix in version 2026.2.22 tightens identity matching to sender-scoped fields only and requires explicit prefixes for mutable metadata. This vulnerability is considered an authorization hardening issue in the context of 1:1 or trusted group chats.
OpenClaw CLI plugins run in-process with the same OS privileges as the main process, and runtime command execution helpers are part of the trusted plugin boundary. This advisory clarifies that these helpers do not introduce a new sandbox boundary. Users should only install trusted plugins and use explicit plugin ID pinning. A fix is planned in version 2026.2.19.