Goldman Sachs bans traders and bankers from AI coding tools

In October, Goldman Sachs updated its internal guidance and restricted access to Anthropic's Claude Code and Microsoft's GitHub exclusively to engineers and quantitative strategists — ending a months-long experiment that allowed bankers and traders to build AI-powered applications for their own work tasks.

The decision, first reported by Bloomberg on November 2, follows a trial period during which the bank assessed productivity gains against security and compliance risks. According to spokesperson Nick Carcaterra, Goldman "completed an experimentation period" and is now placing the tools "in the hands of those best suited to leverage them for the long term."

The end of 'vibe coding' at Goldman

As recently as July 2026, Chief Information Officer Marco Argenti had praised "vibe coding" — the practice of non-technical employees using generative AI to prototype solutions that would then be refined by the bank's 12,000-person engineering workforce. The model promised to close the gap between spotting a business problem and shipping a working fix.

Traders had used Claude Code to monitor client trading flows, model positions, and produce written and graphic materials. Now, access has been cut. Some staff expressed frustration with the change, according to Bloomberg.

What shifted was the category of employee — not the geography. A prior restriction in April had targeted Hong Kong staff; the new guidance is functional, separating those who write code from those who trade with it.

Why Goldman pulled back

Goldman has not disclosed any specific security incident or quantified the findings behind its decision. But the risks troubling the sector are well understood: an application can produce plausible numbers while applying wrong pricing assumptions, and a coding agent with access to internal systems can pull confidential data, link restricted databases, or push changes outside the bank's approval process.

Financial software cannot be judged solely by whether it runs — it must work correctly under compliance rules, audit documentation, and data governance that have no place in personal productivity prototypes.

What happens to those who lost access?

Goldman is not backing away from AI broadly. It is rolling out its proprietary GS AI Assistant to approximately 46,000 employees, combining models from OpenAI, Anthropic, and Alphabet. Staff outside engineering who relied on Claude Code or GitHub have been directed to request alternatives or use that internal assistant.

The bank is also running agentic AI trials, including in trade surveillance — suggesting the shift is about which tools sit in which hands, not whether Goldman uses AI at all.

A mirror for the financial sector

Goldman's decision fits a pattern of growing caution. In July 2026, Citigroup also drew a clear line: it does not allow third-party agents to log into accounts or execute trades. JPMorgan Chase, in July, instructed employees to select AI models based on task complexity rather than using expensive models for routine activities.

By contrast, BNY Mellon allows non-technical employees to build on its Eliza platform, within an environment designed around the bank's data, legal, and compliance requirements. BNY lists more than 400 AI-enabled solutions in production, and AI now writes approximately 40% of its code, PYMNTS reported in April.

Deutsche Bank keeps the tools with professional developers: over 15,000 of them use GitHub Copilot and Google Gemini Code Assist, saving an estimated 1.5 to 2.5 hours per week, according to the bank itself.

MUFG built an in-house alternative to Claude Code after discovering the tool could work around internal controls, and maintains a human-in-the-loop on all its AI tools.

What to watch

The tension between productivity and control is not new in investment banking — but "vibe coding" brought an unprecedented variable: the ability to create working software without formal software engineering training. The problem is that functional financial software requires compliance validation, audit documentation, and data governance that go far beyond business logic.

Code agents are incredible for prototyping — and perhaps that is exactly why no one should be using them in production. The question is whether Goldman is making a temporary correction, adjusting the boundaries of its experimental phase, or signaling the end of a trend that other banks are still exploring. In an industry where the price of a mistake is measured in billions, the answer will likely always be the same: AI has its place, but not in any form that anyone can use.

Sources: Bloomberg, PYMNTS, Crypto Briefing

✓ Independent sources cross-checked and verified before publishing