Solutions  ·  2026-06-20

AutoJack: RCE Exploit Chain via Localhost Trust Boundary in AI Browsing Agents (AutoGen Studio)

SolutionsHigh impactGlobal
Microsoft disclosed AutoJack (June 18, 2026) — an exploit chain allowing untrusted web content rendered by an AI browsing agent to reach a local MCP WebSocket and execute arbitrary processes on the host. Chains three weaknesses: localhost origin bypass (agent inherits local identity), missing authentication on MCP paths, and unsafe parameter injection via URL. Vulnerable code never shipped in PyPI releases; upstream hardened before publication.
Demonstrates systemic risk pattern: localhost trust boundary becomes attack surface when agents can browse the web and access local services simultaneously. Converts browsing agent into delivery vehicle for RCE; pattern extends beyond AutoGen to any framework allowing agent web browsing + local tool access.
Developers using AutoGen Studio or similar frameworks with web-browsing and local service capabilities; immediate audit required for localhost-bound control planes exposed to agent browsing.
Sources
Microsoft Security Blog: AutoJack: How a single page can RCE the host running your AI agentCSO Online: Microsoft says web-enabled AI agents can trigger host-level RCE
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →