Kyvvu

Agent incident database

Semantic Kernel: prompt injection could turn an agent tool into host code execution

6 Feb 2026Microsoft · Semantic Kernel (.NET SDK)unauthorised code executiondisclosed vulnerability

A file-download function in Microsoft's Semantic Kernel agent framework was accidentally exposed to the AI model as a tool, letting it write files anywhere on the host machine. Microsoft showed how injected instructions could use this to plant a script that runs at the next login.

How it unfolded

Each step on its own looked like ordinary agent work. The harm came from the sequence.

  1. 1

    Read content containing injected instructions

  2. 2

    Model chose to call the agent's code tools as instructed

  3. 3

    Ran code in the sandbox that created a malicious script

  4. 4

    Copied the script from the sandbox to the host's Windows Startup folder

    harm
  5. 5

    Script ran on the host at the next user sign-in

    harm

Which policy would have stopped this?

In our reading of the reports, these policies from the open library stop the harmful step before it runs.

Open Code execution requires a preceding gate in the playground and build the path yourself.

This is our reading of public reporting. We haven't run it against the original systems.

What happened

CVE-2026-25592, published on 6 February 2026, affects the SessionsPythonPlugin in the Semantic Kernel .NET SDK before version 1.71.0. The DownloadFileAsync method had been marked with a [KernelFunction] attribute, which advertised it to the model as a callable tool with a model-controlled localFilePath parameter and no path validation. In the attack chain Microsoft described on 7 May 2026, one injected prompt makes the agent use ExecuteCode to create a malicious script inside the code sandbox. A second makes it call DownloadFileAsync to copy that script to the host's Windows Startup folder, where it runs at the next user sign-in. Microsoft disclosed a related flaw, CVE-2026-26030, in which the Python SDK's in-memory vector store built a filter with eval() from model-controlled input.

Outcome

Microsoft fixed it in Microsoft.SemanticKernel.Core 1.71.0 by removing the attribute so the model can no longer call the function, and added path validation; the related Python flaw is fixed in semantic-kernel 1.39.4. Microsoft's post does not say whether either was exploited in the wild.

Sources

A summary of public reporting. The step-by-step timeline is our reading of those sources; where they differ, we say so above.

Your agents

Could this happen to yours?

For security teams

Request a demo

We walk through your agents' riskiest paths with you.

For developers

Build a path yourself

Recreate a sequence in the playground and see how policies treat it.