Security research, datasets, and reproducible findings from AgentGuard across AI agents, plugins, MCP ecosystems, and agent runtimes.
AgentGuard Research publishes ecosystem-scale security studies with supporting datasets, methodology, and verification information.
September 2, 2026
AgentGuard sampled 659 DeepSeek Harness plugins representing 90.0% of marketplace downloads over the previous 30 days.
Among 590 successfully scanned plugins, 94.7% were rated High or Critical based on capability surface, permission exposure, and potential impact.
Study resources:
- Research methodology and findings
- Complete 659-target dataset
- Scan status and capability results
- Commit and artifact verification information
- Dataset SHA-256 checksum
AgentGuard is a security review tool for AI agents and plugin ecosystems.
It helps users, developers, and marketplaces inspect capability surfaces, permissions, reachable resources, and high-impact code paths before code is granted access to an agent runtime.
Secure AI before it acts.
Powered by AgentGuard.