Goldman Sachs Bars Traders and Bankers From Using Claude Code and GitHub Copilot

Goldman Sachs has pulled AI coding tools — including Anthropic's Claude Code and GitHub Copilot — from traders and bankers, restricting access to roughly 12,000 engineers after a productivity and risk review found the bank-wide experiment created compliance exposure. The reversal is one of Wall Street's clearest signals yet that early enthusiasm for agentic AI tools is colliding hard with financial sector risk controls.

Bloomberg Law reported that the restriction affects Goldman employees in trading desks and investment banking divisions who had been using the tools for tasks including market simulation, client-flow analysis, and drafting client communications. Those use cases are now locked out — access is limited to engineers in the technology division.

The Reversal

Goldman CIO Marco Argenti publicly endorsed what he called "vibe coding" — using AI tools to generate functional code without deep programming expertise — in July 2026, framing it as a productivity unlock for non-technical roles across the firm. The bank-wide rollout followed.

The reversal less than four months later points to a gap between productivity gains in controlled settings and the risk profile of the same tools operating in a heavily regulated, market-sensitive environment.

The specific tools restricted include:

• Anthropic's Claude Code (an AI-powered coding agent with agentic capabilities)
• GitHub Copilot (Microsoft's AI pair-programming tool, widely used across industries)

Both tools are designed to write, edit, and execute code. In the hands of engineers building internal systems, that's well-contained. In the hands of traders using AI to write scripts that interact with market data or client communications systems, the risk surface is fundamentally different.

Why Compliance and Model-Risk Frameworks Object

Financial institutions operate under a layered set of regulations that treat automated decision-making and market-connected code with particular scrutiny. Goldman's restrictions likely reflect concerns in several areas:

Model risk management (MRM). U.S. banking regulators (OCC, Fed, FDIC) require that any model used in a material business decision be validated, documented, and governed. An AI coding tool that generates scripts for market simulation or flow analysis may generate outputs that function as undocumented models — triggering MRM requirements the bank isn't yet equipped to satisfy at scale.

Data exfiltration risk. Coding tools that send prompts to external APIs (even Anthropic's or GitHub's) create potential channels for proprietary trading data, client information, or market positioning to leave the firm's controlled environment. That's a compliance and competitive concern regardless of the tool's stated privacy policies.

Auditability. Regulators expect banks to be able to explain how consequential decisions were made. Code generated by an AI agent may not produce the kind of documentation trail that satisfies audit requirements.

The Broader Wall Street Signal

Goldman isn't alone. Earlier this year, several major banks — including JPMorgan and Barclays — established formal review processes for AI tool adoption, with trading-desk access treated as a higher-risk category. Goldman's restriction formalizes what has been a growing informal consensus: the productivity case for AI coding tools is strongest for software engineers, and weaker for roles where the outputs interact with markets, clients, or regulated processes.

The development is an uncomfortable data point for Anthropic and Microsoft (GitHub's parent), both of which have been aggressively marketing their tools to financial services. Goldman's adoption had been cited in industry discussions as evidence that even the most risk-averse financial institutions were coming around to agentic coding tools.

For enterprises evaluating AI tool rollouts, Goldman's experience suggests a useful framework: the risk profile of a coding tool isn't determined by the tool itself, but by the downstream systems and data its outputs interact with. Engineers building internal tooling carry different risk than analysts running market simulations.

What to Watch

Goldman's move may accelerate formal regulatory guidance on AI coding tools in finance — an area where U.S. banking regulators have been notably quiet compared to their guidance on AI in credit decisions. Watch for OCC or Fed guidance on "AI-generated code" as a subcategory of model risk, and for other major banks to quietly follow Goldman's segmentation playbook.

By Hector Herrera