🔥 Trending on HN

Goldman Sachs Limits Claude Code at the Trading Desk

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI coding agent(AI coding agent)

A program that can write and run computer code from instructions.

information barrier(information barrier)

A bank rule that keeps sensitive information apart.

strats(strats)

Employees who combine trading, research, and software work.

State of AI, an AI news site, reported on Nov. 2 that Goldman Sachs, a major US investment bank, narrowed access to Anthropic’s Claude Code and Microsoft’s GitHub coding tools. The report cites Bloomberg News and internal guidance updated in October. Most bankers and traders can no longer use the tools at work. Engineers remain the main permitted users. Employees who combine trading, quantitative research, and software engineering, known as strats, reportedly remain outside the restriction.

What changed

Claude Code is an AI coding agent. It can read files, write programs, and run them, normally with the user’s access. The reported rule targets third-party agents that can create and execute code. Goldman’s internal assistant, which uses models from several companies, remains available across the firm. This is not a report that Goldman banned all AI or all use of Claude. A Goldman spokesperson said the bank makes some AI tools available while it determines their best uses.

The background

In July, CIO Marco Argenti had praised a workflow in which nontechnical employees made first versions of tools. Engineers then improved them. Goldman had described more than 12,000 developers among roughly 47,000 staff. The October guidance shifts access and first-line control back toward engineering. Goldman had also removed Claude from its internal platform for Hong Kong bankers in April. That earlier move concerned location and contract terms. The new restriction is based mainly on job role. Goldman still works with Anthropic engineers on Claude-based agents for trade accounting and client onboarding. Engineers operate those agents.

Why the rule matters

An AI coding agent can make a useful tool quickly. It can also create a control problem. A trader’s script might monitor client flows or model positions. Banks use information barriers to keep sensitive deal and trading information apart. If an agent-written script enters daily work without review, testing, or deployment controls, unusual errors may go unnoticed. The report also mentions productivity concerns and token-based costs. It does not show how much each factor influenced the decision.

What is confirmed

Bloomberg reported that it had seen the October guidance. State of AI also reported that some strats could keep access. The story drew 616 points and 361 comments on Hacker News, an online technology forum. Those figures measure community attention. They do not prove the report is correct.

What remains unclear

The public reporting does not say how many workers lost access, which exact functions are blocked, or how long the rule will last. Goldman has not published the productivity or risk data behind the review. Nothing in the report establishes a data leak or trading loss. It also does not describe a complete break with Anthropic.

What to watch next

The next question is whether Goldman creates a controlled path for nonengineers. That could include a named reviewer, clear data limits, and testing before a tool reaches important work. The source article argues for that kind of ownership, but Goldman has not announced such a replacement. Watch whether other banks adopt role-based limits, and whether Goldman later expands or narrows access. The larger issue is accountability: who checks AI-written code, and who answers when it touches real financial work?

💬 HN debates Goldman Sachs restricting Claude Code

The thread questioned both the article’s provenance and the policy itself. The central dispute was whether coding agents on finance workers’ machines are too risky, or whether a constrained platform could let non-engineers build useful tools safely.

  • Some commenters alleged that the article repackages reporting from Bloomberg or resembles an AI-written site used to promote a company. That is a commenter claim, not a fact independently established in the thread.
  • Several commenters focused on the combination of client and trading data, email, internal networks, and internet access on a banker’s or trader’s workstation. A semi-unattended agent could turn a small mistake or prompt injection into a serious leak.
  • Commenters’ self-reports from receiving finance-built tools said the visible interface is often the easy part. Architecture, database schema, security, edge cases, source control, deployment, and maintenance can remain unresolved, leaving engineers to rewrite the tool or leave it in a backlog.
  • A strong counterproposal was to build a controlled programming environment with least-privilege APIs, sandboxes, deterministic tests, linters, analyzers, and audit logs. Engineers would build the rails, while bankers would use them within defined limits.
  • The opposing response was that enough safety rails eventually recreate a set of fixed, approved applications. If users cannot understand or inspect generated code, the claimed flexibility may be mostly an illusion, and the organization still depends on platform maintainers.
  • Data governance was treated as distinct from sandboxing. A managed-agent system might enumerate permitted sources and prevent customer complaints or deal information from reaching unrelated workflows, but provider retention, mistaken internal permissions, and accidental crossings of information barriers would remain risks.
  • Concrete failure scenarios included an app that creates an unmonitored communication channel across an information barrier, or one that summarizes privileged email into a database accessible to a trading desk. These were presented as compliance and liability problems, not merely arguments against nontechnical users.
  • Another counterargument challenged the assumption that engineers are a sufficient safety boundary: if developers also stop reading large volumes of generated code, human review becomes unreliable. Supporters of more agentic workflows argued for standardized architecture and machine-checkable evidence; critics said it remains wishful thinking rather than a proven solution.

growing digest at 455 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Why Did Goldman Sachs Restrict AI Coding Tools?

📰 Full story: Goldman Sachs Limits Claude Code at the Trading Desk

Goldman Sachs now limits who may use tools that write computer programs.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
information barrier(information barrier)

A rule that keeps important information between teams.

strats(strats)

Workers who combine trading, research, and software work.

💡 The gist

  • Goldman limits Claude Code and GitHub coding tools.
  • Most bankers and traders cannot use them at work.
  • Engineers can still use them, because they can review the code.

Goldman Sachs is a large investment bank. Anthropic is the company that makes Claude. Claude Code is an AI tool for writing computer programs. It can read files, write code, and run it. Microsoft, a software company, also has GitHub coding tools. Goldman restricted those tools too.

State of AI, an AI news site, reported the change on Nov. 2. It cited Bloomberg News and October internal guidance. Most bankers and traders lost access at work. Engineers kept access. Some strats kept access too. Strats combine trading, research, and software work.

Why does this matter? The tool normally uses the worker’s access. It may reach files that worker can reach. A trader might use it to watch client trades. A trader might model market moves or positions. Those tasks can involve sensitive information.

Banks use information barriers. These rules keep some information between teams. For example, deal information should not reach a trading team. A quick program may cross a boundary by mistake. An engineer can review and test the program first. That helps catch errors. It can also show who owns the tool.

Goldman had supported a different plan in July. Workers could build a first version with AI. Engineers then improved the tool. The October rule puts engineers in control earlier. Goldman has about 12,000 developers. It has about 47,000 staff, according to the report.

This does not mean Goldman stopped using AI. Its internal assistant still uses several models. Goldman also works with Anthropic on office agents. Engineers run those agents. In April, Goldman limited Claude for Hong Kong workers. That case involved location and contract terms. The new case mainly involves job roles.

Hacker News, a technology discussion site, showed strong interest. The link received 616 points and 361 comments. Those numbers show attention from that community. They do not prove the report is true.

Goldman has not published all review data. We do not know how many workers lost access. We do not know how long the rule will last. Next, watch for a safe path for nonengineers. Also watch whether other banks follow Goldman’s move.

💬 Why Goldman may limit Claude Code

HN commenters questioned the article’s source and disagreed about who should use coding AI at a bank. Some want it limited to engineers; others want a safer system that lets more people use it.

  • Some commenters suspected the article was copied from another report or came from an AI-written promotional site. That suspicion was not independently verified in the thread.
  • Bankers’ computers may contain customer data, trade information, email, and access to company systems. An AI agent can make mistakes or follow a malicious instruction, so one small error could expose sensitive information.
  • Commenters’ self-reports say that the hard parts of finance-made tools are usually security, databases, testing, deployment, and maintenance—not making a screen that appears to work.
  • Some commenters want a locked-down platform with small permissions, sandboxes, approved building blocks, tests, and audit logs. Others say enough restrictions turn it back into a set of approved programs.
  • A managed AI service might control which data a workflow can read, but it would not automatically solve provider access, mistaken permissions, or breaches of internal information barriers.
  • Even engineers may not safely review huge amounts of generated code. The thread therefore argued about combining human review with automatic tests and standard designs.

growing digest at 455 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Why Did a Bank Let Fewer People Use Its Computer Helper?

📰 Full story: Goldman Sachs Limits Claude Code at the Trading Desk

Goldman Sachs now lets fewer workers use Claude Code.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Claude Code(Claude Code)

A computer helper that writes and runs instructions.

engineer(engineer)

A person who builds and checks computer tools.

Goldman Sachs is a big bank. Anthropic is the company that makes Claude. Claude Code is a computer helper. It can read files. It can write computer instructions. It can run those instructions. A bank has important files. So Goldman now limits who uses the helper. Most bankers and traders cannot use it. Engineers can still use it. Some workers who mix trading and computer work can too. They can check the computer instructions more carefully. Goldman did not turn off all AI. Its own helper still works. Engineers also build office tools with Anthropic. In April, Goldman stopped Claude for some Hong Kong workers. That was a different rule. Hacker News is a place where people discuss technology. This story got 616 points and 361 comments there. That means many people noticed and discussed it. It does not prove the story is true. We still do not know how many workers lost access. The big question is simple. Who checks the computer instructions?

💬 Why a bank might limit its coding helper

Some people worried that the news story itself might not be fully trustworthy. They also worried about letting a computer helper work with bank secrets.

  • The helper might see secret customer or trading information. A mistake could let the wrong person see it.
  • Commenters’ self-reports say a quick little app may look finished but still be unsafe or hard to fix.
  • Some people want to put the helper in a safe box. Others say a box with too many rules is just an ordinary approved app.
  • People must check the helper’s work, because even programmers can miss mistakes.

growing digest at 455 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources