Goldman Sachs is pulling Claude Code from its bankers and traders
Goldman Sachs is restricting Anthropic’s Claude Code and GitHub Copilot to software engineers, blocking their use by traders and bankers over security, compliance, and operational risk concerns. Commenters highlight dangers of “coding agents” running on trading-floor machines with access to sensitive data and production systems, invoking issues like prompt injection, data exfiltration, insider‑trading exposure, and unreviewed “vibe‑coded” tools becoming critical infrastructure. The move is seen as part of a broader struggle inside large firms to capture AI productivity gains while avoiding opaque, hard‑to‑govern systems—especially awkward given Goldman’s role in Anthropic’s upcoming IPO.
Scope of Goldman’s Change
- Internal guidance now restricts Claude Code and GitHub Copilot to engineers; traders/bankers lose direct access.
- The firm’s own in‑house AI assistant (non‑coding) remains widely available.
- Some commenters note earlier similar restrictions by region (e.g., Hong Kong desks).
Risk, Compliance & “Lethal Trifecta”
- Major concern: coding agents on primary trading/banking machines combine three risks: access to sensitive data, ability to execute code, and network connectivity.
- People highlight leak risks (client data, internal models, deal flow), insider trading concerns, and unmonitored cross‑wall communication between banking and trading.
- Several suspect an operational incident (bad trade, data leak, or compliance breach) likely triggered the clampdown, not token cost.
Non‑Engineers Coding & Productionization
- Longstanding pattern: front‑office staff creating fragile tools (spreadsheets, macros, now AI‑generated apps) and pushing engineers to “productionize” them.
- Engineers report such tools are often unusable beyond rough prototypes and can become unsupported critical-path systems.
- Some think letting non‑technical staff build apps without review is itself a compliance issue.
Agentic Engineering & “Vibe Coding” Debate
- One camp argues you must go “all in” on agentic engineering: strong patterns, guardrails, linters, architecture, and a managed “enterprise harness” so almost anyone can safely build.
- Skeptics call this wishful thinking; say agents are bad at enforcement, prompt injection remains unsolved, and unreviewed AI code is inherently risky.
Security Architectures & Sandbox Proposals
- Many argue agents should run in tightly sandboxed VMs, remote workspaces, or internal “AI gateways” with strong federation, access control, and DLP—not on traders’ desktops.
- Others note sandboxing is hard in practice; corporate policies often disable virtualization, and even sandboxes have been escaped.
Cost, Governance & Broader AI Adoption
- Some organizations are cutting or shifting AI usage due to unpredictable API billing and weak enterprise cost controls.
- Commenters expect a wave of AI governance products: centralized platforms, MCP/connector layers, and policy‑enforced “citizen developer” environments.
- Views diverge on impact: some see this as prudent risk management in finance; others as a step back that could leave traders less competitive.