Critical “Log Poisoning” Vulnerability in OpenClaw AI Agent Allows Malicious Content Injection
OpenClaw, an open-source AI assistant, has addressed a "log poisoning" vulnerability that allowed remote attackers to inject user-controlled content into logs. This issue is detailed in the OpenClaw Security Advisory and affects versions prior to…
OpenClaw, an open-source AI assistant, has addressed a "log poisoning" vulnerability that allowed remote attackers to inject user-controlled content into logs. This issue is detailed in the OpenClaw Security Advisory and affects versions prior to 2026.2.13.
The vulnerability involves an indirect prompt-injection attack where untrusted input is recorded in logs that an AI agent may later use for troubleshooting, potentially affecting the agent's reasoning process. The vulnerability was related to the logging of certain WebSocket request headers, such as Origin and User-Agent , without proper sanitization in affected versions.
If an attacker accessed an OpenClaw gateway interface, they could embed crafted header values in the logs, creating a "poisoned" log trail. This could impact workflows where logs are used for diagnostic purposes, as the agent might misinterpret the data.
The practical impact of the vulnerability is contingent on how logs are utilized, especially when the agent uses them to diagnose errors. Injected content in logs could be misconstrued as legitimate input, potentially altering troubleshooting processes or decisions.
This issue is detailed in the OpenClaw Security Advisory and affects versions prior to 2026.2.13.
Instances of OpenClaw exposed on the internet, particularly via default port 18789, represent an increased attack surface.
The issue was resolved in OpenClaw version 2026.2.13. Users are advised to upgrade to this version or later and ensure that gateway interfaces are secured with strong access controls to prevent unauthorized internet access.
It is recommended to treat agent-consumable logs as untrusted inputs. Implementing measures such as sanitizing or redacting user-controlled fields, capping header sizes, and separating logs for human debugging from agent inputs can mitigate risks. Monitoring for unusual patterns in headers and failed WebSocket connections is also advised as they may indicate attempted log poisoning.
Based on reporting by Cyber Security News.
