Enterprise AI Agents Create Inventory, Identity and Governance Gaps, Island Says

· Original article ↗

Summary

Island argues that enterprise AI agents create gaps in inventory, identity, attribution and oversight. Its research reports security findings in MCP servers and malicious repositories, and recommends real-time governance and just-in-time access.

Key points

  • Island reports scanning 33,563 published MCP server builds containing 475,865 tools; nearly half had at least one security finding.
  • Island says about one in eight scanned builds exposed a tool capable of code execution, data deletion or another irreversible action on its first call.
  • The article says agent activity can be difficult to distinguish from human actions in logs, complicating attribution and audit trails.
  • Island's team reports finding roughly 7,600 malicious GitHub repositories, including more than 800 posing as AI Skills or MCP servers, in a campaign delivering the StealC information stealer.
  • The article recommends live inventories of agents and their capabilities, distinct agent identities, and just-in-time access scoped to individual tasks.
  • It advocates inline checks of prompts, tool calls and returned data, plus usage and cost monitoring based on actual sessions and tool calls.

Article Details

Topic
Enterprise AI agent security governance, including agent inventory, identity attribution, inline controls, and malicious repository discovery.

Malware

Vendors

Industries

Related Articles