LayerX Report: AI Extensions - The Unmonitored Consumption Channel

Hey team,

Saw the LayerX report drop today regarding AI browser extensions acting as a silent consumption channel. We spend so much time blocking the ChatGPT UI at the proxy level that we've ignored the backdoor sitting right in the user's browser.

The report highlights that these extensions aren't just risky due to malicious code; they're risky by design. "Helper" extensions that summarize emails or web pages are scraping sensitive context and sending it to third-party LLM endpoints, completely bypassing our DLP controls. Since the traffic is browser-initiated to legitimate API domains (like OpenAI or Anthropic), standard egress filtering often misses it.

I've started auditing our fleet for extensions with overly broad permissions. Here's a quick PowerShell snippet to enumerate installed extensions for Chrome to cross-reference with your allow-list:

$extensi
if (Test-Path $extensionsPath) {
    Get-ChildItem $extensionsPath -Directory | ForEach-Object {
        $manifest = Join-Path $_.FullName "*\manifest."
        if (Test-Path $manifest) {
            $data = Get-Content $manifest | ConvertFrom-Json
            [PSCustomObject]@{
                Name = $data.name
                Version = $data.version
                Permissions = $data.permissions -join ', '
            }
        }
    }
}

The scary part is that most of this traffic looks like standard API calls. How are you guys handling this? Are you straight-up blocking all extensions, or have you found a way to inspect the LLM-bound payloads?