🔗 Securing Claude Cowork: a security practitioner’s guide

Claude Cowork is not just another chatbot; it presents unique security challenges since it runs code locally and executes tasks independently. Prompt injection is the most significant risk, with a roughly 1% attack success rate reported by Anthropic even after mitigations. Using OpenTelemetry can enhance visibility, though its limitations mean it should be integrated with SIEM systems for real-time monitoring. Organizations should also adopt a suitable deployment posture, from lockdown to open enablement, depending on their risk tolerance. Importantly, understanding the defaults and carefully managing access to features helps reduce potential vulnerabilities.