Enterprise AI Coding Agent Security: Why Goldman Sachs Limits Access

The enterprise question around AI coding agents is shifting from who can use them to what they can reach. Goldman Sachs reportedly limits Anthropic’s Claude Code and Microsoft’s GitHub coding tools to engineers, while allowing some quantitative employees, known as strats, to keep access. The policy points to a broader tension: agents can speed up work, but running them on employee machines may expose credentials and internal systems.
Replit CEO Amjad Masad described endpoint-based agents as a systemic security risk for companies and governments in a Tuesday post on X. He said financial firms are restricting tools such as Claude Code, adding that “This is widespread in finance. The ban hammer is coming down.” Masad attributed the concern in part to supply chain attacks.
Access is only part of the control problem
Goldman spokesperson Nick Carcaterra told Bloomberg that the bank weighs potential productivity gains against risks when deciding which employees can use AI tools. That approach draws a line between staff groups, but limiting access to engineers alone doesn’t contain every threat. Developers can still encounter malicious instructions or compromised software while working with agents.
The risks include agents reading credentials stored on employee devices, acting on hostile instructions, or pulling in compromised dependencies. Vendors have been moving toward short-lived credentials to reduce the value of keys that persist on a machine. Vercel, for example, recently gave suppliers a deadline to replace long-lived keys that an endpoint agent could potentially read and reuse.
Asked whether banks are more concerned about code leaving their environments or weak audit trails, Masad replied, “So many things.” He didn’t cite specific incidents behind Goldman’s reported restrictions or provide evidence showing how many financial institutions have adopted similar policies.
Replit’s case for controlled cloud environments
Masad shared a security advisory published by Replit on Monday. The document, written by Replit CTO Luis Héctor Chávez and engineering lead Manish Mehta, recommends running agents in company-controlled cloud environments instead of directly on employee devices.
“AI agents should operate in dedicated, organization-controlled environments with limited access to the resources each task requires,”
The proposal puts the agent in an isolated workspace and keeps credentials, audit logs and approvals under separate management. Its aim is to constrain an agent’s reach for each task, rather than rely only on deciding which employees may run it.
Replit’s advisory cites a compromised Claude Code extension distributed through the Cursor marketplace as an example of attackers exploiting developer tools. Other incidents show how a compromise can spread through connected accounts and repositories: attackers hijacked roughly 1,900 GitHub accounts and then targeted about 30,000 repositories.
Even tightly managed access can’t guarantee that an agent will stay within scope. A separate report on coding agents describes how they can follow malicious instructions or bring in compromised software. Research cited in the source found that removing a single line from Claude Code’s instructions increased out-of-scope actions. For enterprise teams, the practical challenge is controlling agent permissions and behavior alongside user access.
A deployment decision, not just a user policy
Goldman’s reported limits suggest that some companies are reconsidering broad employee access to coding agents. These tools can let non-developers build software, but that promise becomes harder to balance when an agent can reach sensitive corporate systems from a personal or work device.
Masad’s position is that companies can keep using agents if they move them off employee endpoints and restrict their access. Whether financial firms broadly adopt this model remains unclear: his description of restrictions as widespread came without public evidence on their scale.
The New Stack said it had contacted Replit for further comment on the reported restrictions, the security concerns involved and whether financial institutions are changing their approach to AI coding agents.
The post Goldman Sachs restricted Claude Code for bankers — but made one surprising exception appeared first on The New Stack.