Clawed and Dangerous: Can We Trust Open Agentic Systems?

The researchers systematically review the security issues of open agentic systems from a software engineering perspective, propose a six-dimensional analytical taxonomy, and synthesize 50 papers covering attacks, benchmarks, defenses, auditing, and engineering foundations; OpenClaw is a typical instance of this class of systems discussed in the paper. The authors note that the plans of such systems are generated at runtime and that key decisions may be influenced by untrusted natural language inputs and tool outputs; the core challenge is governing agent behavior under persistent uncertainty rather than merely defending against single attacks. The survey provides reference guidelines for secure-by-construction agent platforms and a scorecard for assessing platform security posture, and finds that existing literature is relatively mature in attack characterization and benchmark construction but remains weak in deployment control, operational governance, persistent memory integrity, and capability revocation; these gaps constitute an engineering agenda for building a governable, auditable agent ecosystem that remains resilient after compromise.