OpenWorker Open-Source Security Agent

Andrew Ng announced a new version of OpenWorker, an open-source AI agent framework that has been repurposed by users for security applications. The update includes built-in cybersecurity agents. The framework's key architectural concept is the separation of model and harness: the model handles reasoning while the harness manages tool scheduling, command execution, file operations, and data routing. The harness is fully open-source and auditable, addressing concerns about code exfiltration. Three built-in security capabilities are included: code vulnerability scanning (driving existing scanners like semgrep, then using AI to triage by reachability and generate fixes), dependency supply chain injection detection (targeting malicious packages, typosquatting, and compromised CI workflows), and cloud security configuration checks (scanning infrastructure-as-code and verifying against live environments). The framework is model-agnostic, supporting ChatGPT subscriptions, API keys, or fully local open-weight models, which is critical for security work that may trigger refusal from closed-source models.
Key insights
The Model-Harness Split Is Critical for Security and Trust
The key architectural insight is separating the model (the reasoning brain) from the harness (the software layer that executes actions). Data exfiltration risk primarily resides in the harness, not the model. By making the harness fully open-source and auditable, OpenWorker allows security teams to verify there are no backdoors. This contrasts with closed-source coding agents that require unconditional trust in the vendor.
Open-Weight Models Are Necessary for Legitimate Security Work
Legitimate defensive security work—such as reproducing known vulnerabilities for remediation, penetration testing, and malware analysis—often triggers refusal from closed-source models due to alignment policies. Local execution of open-weight models removes this restriction and ensures sensitive code never leaves the machine. This is a hard requirement for financial, government, and other regulated teams.
AI Agents Can Shift Security Left by Generating Fixes, Not Just Findings
OpenWorker's security agents don't just identify vulnerabilities; they triage findings by reachability (whether vulnerable code is actually executable) rather than severity alone, and generate fixes as pull requests in the codebase's style. This 'shift left' approach places security remediation directly in the development workflow, where fixes are cheapest to implement.