Four vulnerabilities in the OpenClaw AI assistant can be chained together to plant backdoors on the underlying host, cybersecurity firm Cyera warns.
The bugs, collectively known as Claw Chain, allow an attacker with code execution privileges inside the sandbox to control the agent runtime and abuse it to compromise the system.
According to Cyera, the attacker can rely on prompt injections, malicious plugins, and compromised external input to trigger the attack chain and turn the AI into their own assistant.
After gaining code execution within the OpenShell sandbox, the attacker can exploit a race condition (CVE-2026-44113) to read files outside the mount root, or an exec allowlist analysis bug (CVE-2026-44115) to execute unapproved commands at runtime.
Successful exploitation of these issues, Cyera notes, allows the attacker to bypass sandbox restrictions and leak credentials, API keys, tokens, configuration files, and other sensitive data.
Next, the attacker can exploit an MCP loopback flaw (CVE-2026-44118) to manipulate the unverified ownership flag and elevate their privileges to owner-level. The attacker gains access to critical management functions, including configuration and orchestration of execution.
Finally, the attacker can exploit the fourth vulnerability, a critical-severity race condition in the OpenShell sandbox (CVE-2026-44112, CVSS score of 9.6), to write data outside the sandbox boundary. It allows the attacker to modify configurations, plant backdoors, and gain persistent control of the host.
“By weaponizing the agent’s own privileges, an adversary moves through data access, privilege escalation, and persistence – using the agent as their hands inside the environment. Each step looks like normal agent behavior to traditional controls, broadening blast radius and making detection significantly harder,” Cyera says.
The cybersecurity firm says there are over 60,000 publicly accessible OpenClaw instances, noting that the agents typically have broad access to internal systems, sensitive data, and secrets.
Attackers successfully chaining the Claw Chain bugs could compromise environment variables, tokens, authentication material, internal configurations, system credentials, source code, user prompts and outputs, conversation history, and privileged operations.
“Importantly, this chain does not rely on a single critical exploit like arbitrary command execution. Instead, it demonstrates how multiple smaller weaknesses (data leakage, race conditions, and improper access control) can be exploited in parallel from a single foothold to achieve a full compromise scenario,” Cyera notes.
All four vulnerabilities were reported to OpenClaw’s maintainers on April 22, and patches were rolled out the next day.
Related: Mythos Proves Potent in Vulnerability Discovery, Less Convincing Elsewhere
Related: AI Coding Agents Could Fuel Next Supply Chain Crisis
Related: AI Fuels ‘Industrial’ Cybercrime as Time-to-Exploit Shrinks to Hours
Related: Why Agentic AI Systems Need Better Governance – Lessons from OpenClaw
