Goldman Sachs has pulled its Goldman Sachs AI coding tools access from most non-engineers. According to The Decoder, relaying a Bloomberg report published November 2, 2026, traders and bankers no longer get Claude Code or GitHub Copilot.
Engineers keep both tools. The bank says the experiment is over, and it wants these tools in the hands of people who can use them long term.
Here are the facts that matter most.
- Tools affected: Anthropic’s Claude Code and Microsoft’s GitHub Copilot.
- Who lost access: most employees outside engineering, including traders and bankers.
- Who keeps access: Goldman’s engineers.
- Reason given: “productivity and risk management,” per a Goldman spokesperson.
- Source: Bloomberg, citing internal guidelines from October.
The number of affected staff hasn’t been reported, so treat any headcount you see with caution.
What Goldman actually changed
Based on the Bloomberg report, Goldman’s internal guidelines from October narrowed who can use the tools. For months, some traders and bankers had used them to monitor trading flows and create content.
That access is gone for most of them. Engineers were not touched.
Spokesperson Nick Carcaterra told Bloomberg the decision reflects “a variety of factors including productivity and risk management.” He also said the firm is limiting the tools to people “who are best suited to leverage them for the long term.”
From “vibe coding” cheerleading to a rollback
This is a sharp turn. In July, Goldman’s chief information officer, Marco Argenti, publicly praised “vibe coding” by non-technical staff, as The Decoder notes.
The idea was simple. A trader or asset manager would build a quick AI prototype, then hand it to an engineer to make it production-ready. That plan appears to have been dropped.
The same report says the rollback has frustrated some people inside the firm. That’s a reasonable reaction if you built a workflow around the tool.
Why banks get nervous about AI coding tools
Goldman hasn’t spelled out its risk concerns, so anything beyond its two-word reason is our reading, not a confirmed fact. Still, the usual worries are easy to list.
Code written by people who can’t fully review it can hide bugs. In finance, a bug in a script that touches trading data or client information is expensive.
Agentic tools also run commands and read files, which widens the attack surface. We covered one real example in our post on the Claude Code extension backdoor, where a bad build put users at risk.
Here’s how the old and new setups compare, based on what has been reported.
| Group | Before the change | After the change |
|---|---|---|
| Engineers | Claude Code and GitHub Copilot | Still allowed |
| Traders and bankers | Used the tools for flow monitoring and content | Access removed |
| Official stance | Broad experimentation encouraged | Experiment phase over |
The table shows a shift from open experiments to a narrow, controlled rollout. That’s the pattern worth watching across other big firms.
What this means for your own team
You don’t need to work at a bank to learn from this. Many companies handed AI coding tools to everyone, then realized nobody set rules for review.
Agents are already writing a large share of code. Our look at GitHub Octoverse 2026 found AI agents now start over 40% of pull requests, so the review question isn’t going away.
If you manage a team, a few quick checks help. Decide who may use coding agents, and make sure a human engineer reviews anything that ships. Keep these tools away from sensitive data unless your security team has signed off.
What we still don’t know
Plenty is missing from the public report. Goldman hasn’t said how many people lost access, whether any exceptions exist, or what specific risk triggered the review.
We also don’t know whether the bank will approve narrower, supervised uses later. The spokesperson’s wording about “the long term” suggests the door isn’t fully shut, but that’s an inference, not a statement from Goldman.
Until the bank says more, the safest reading is the plain one: the tools stay with engineers, and everyone else loses them for now. We’ll update this post if Goldman or Bloomberg adds detail.
Frequently Asked Questions
Which AI coding tools did Goldman Sachs restrict?
Anthropic’s Claude Code and Microsoft’s GitHub Copilot, according to Bloomberg’s report. Both stay available to engineers.
Who can still use them at Goldman Sachs?
Engineers only. Traders, bankers and most other staff lost access.
Why did Goldman Sachs restrict the tools?
A spokesperson cited “productivity and risk management.” The bank hasn’t given more detail.
Is Goldman Sachs banning AI altogether?
No. The report covers only these coding tools, and engineers still have them. It doesn’t say the bank is dropping AI in general.
How many employees are affected?
That hasn’t been reported. Any specific number you see elsewhere isn’t confirmed in the Bloomberg-based coverage.
Our take
This looks less like a retreat from AI and more like a bank admitting that handing coding agents to everyone was too loose. Expect more firms to land in the same place: engineers get the tools, everyone else gets approved, supervised workflows. If you run a team, set that line yourself before an incident sets it for you.
Leave a Reply