Masad says Goldman Sachs has reportedly restricted Claude Code

Amjad Masad says AI agents running on employee endpoints are becoming a systemic security risk for enterprises and governments, and that some of those organisations are already responding by restricting tools like Claude Code. He names Goldman Sachs as having reportedly done so. Alongside the post he points to a security advisory that, in his words, outlines the risks and recommends a path forward.

Amjad Masad

@amasad

X

AI agents running on employee endpoints are becoming a systemic security risk for enterprises & governments.

That is the whole of the public claim right now. The Goldman detail is reported, not confirmed by the bank, and Masad does not say how wide the restriction goes, when it started, or whether other banks have followed. Nobody has put a number on it yet.

Why a founder should care

The thing being described is not a model safety argument. It is an endpoint argument. A coding agent sits on a corporate laptop with the employee's credentials, their repo access and their shell, and it acts on instructions that can arrive from anywhere it reads. That puts it in the same bucket security teams already use for unmanaged software, which is the bucket where things get blocked by policy rather than debated on merit.

It also follows a pattern that is getting familiar. Masad was the one who said an official Claude Code extension installed backdoors, so his read on agent supply chain risk is not a fresh hobby.

The business read

If big enterprises start treating agents as endpoint risk, the people selling agentic products into them inherit a gatekeeper that did not exist a year ago. Not a procurement call or a security questionnaire, but a laptop policy that says the agent does not run here at all. Founders building agent tooling for large customers should assume the security team gets a veto, and build for the version where the agent has to justify every permission it asks for.