Threat researchCisent
Just connecting an MCP server was enough for an attacker to run code on the machine of anyone using OpenCode
Pillar Security found that a malicious MCP server could take over a Windows machine as soon as authentication began. The maker fixed it without saying it was a security flaw.

On October 19, 2026, Pillar Security published a flaw in OpenCode, an open-source AI coding agent with about 3 million weekly npm downloads and 211,000 GitHub stars. A malicious MCP server could run code on the machine of whoever was using it, with that person’s own permissions.1
What happened
MCP is the protocol AI agents use to connect to external tools and data. To authenticate with a remote MCP server, OpenCode took an address sent by the server itself and handed it to the system to open the browser. On Windows, and in the Windows Subsystem for Linux, a specially crafted address was read as commands. Connecting to that server was enough: the person didn’t have to approve anything or talk to the model.1
The flaw affected versions 1.0.137 through 1.18.32, and was fixed in 1.18.33, released on September 28. macOS and Linux were not affected.1
Pillar reported it privately on August 29. The maker fixed it, but replied that malicious MCP servers were outside its threat model, presented the fix as a change to the code that opens the browser, and did not request a CVE. According to Pillar, the same class of flaw had already received CVEs in three other MCP clients in 2025.1
Why it matters
Without a CVE, vulnerability scanners and dependency alerts don’t flag it. An organization only knows whether it’s exposed if it knows which tools and versions its development team uses.
These agents also work with the developer’s permissions and connect to servers the organization doesn’t control. A server that’s trustworthy today can be compromised tomorrow, or its domain can expire and fall into other hands, as the report itself warns.1
Where Cisent comes in
Cisent Iroh is the agent that trains your team. It talks with people, explains attack techniques and response measures in plain language, and teaches how to respond, for technical and non-technical staff alike: for example, what to check before connecting an MCP server, and why it’s worth updating a tool even without a security advisory. It’s an add-on to any Sentinel product.
Sources
Eilon Cohen, Pillar Security, “Open and shut: remote code execution in opencode via MCP OAuth”, October 19, 2026. https://www.pillar.security/blog/open-and-shut-remote-code-execution-in-opencode-via-mcp-oauth↩︎