Claude Code accelerates, but banks tighten access
Goldman Sachs has restricted access to Claude Code for its bankers and traders, according to Bloomberg, while Replit pushes companies to isolate AI agents in the cloud. Meanwhile, Claude Code’s product lead leaves Anthropic after two years of growing autonomy.
Goldman Sachs has restricted access to Claude Code, Anthropic's coding tool, for bankers and traders, according to Bloomberg, with an exception for certain quantitative profiles. This decision comes as Replit warns about the risks posed by coding agents on employee endpoints, and as Claude Code gains autonomy at Anthropic. The issue matters for finance but also for all companies wanting to deploy AI without opening the door too wide to credentials and code repositories.
Why finance is tightening controls
Replit, whose CEO Amjad Masad spoke on X on November 3, 2026, sums up the problem in a direct phrase: "AI agents running on employee endpoints are becoming a systemic security risk for enterprises & governments." The alert mainly targets tools running directly on employees' machines, where an agent can read access keys, ingest malicious instructions, or carry compromised dependencies. In this context, Goldman Sachs is not acting alone, although the extent of the restrictions remains to be confirmed [to be verified].
Masad speaks of a broad movement in finance: "This is widespread in finance. The ban hammer is coming down," he wrote, without providing public evidence on the number of institutions affected. Bloomberg specifies that the restriction affects bankers and traders, but not certain "strats," those quantitative specialists who design market strategies and models. A way of saying that not everyone has the same right to the copilot, even when they know how to write code.
The real debate: who uses the agent, or where it runs?
The question is not limited to the user's identity. Replit advocates that agents should operate in dedicated environments controlled by the organization, with limited access to the resources necessary for the task. The company also recommends separating credentials, audit logs, and approvals. In other words, the issue is not only to reserve the tool for developers but to take it off the individual workstation to run it within a more sealed perimeter.
This approach aligns with other market signals. Vercel recently required its providers to abandon long-lived keys, which are easier to reuse if an agent or endpoint is compromised. And Replit also cites an incident involving a compromised Claude Code extension distributed via the Cursor marketplace, proof that a toolchain can be weakened very quickly. When one link breaks, the problem often spreads beyond the terminal where it started.
Claude Code changes face at Anthropic
The discussion comes as Claude Code, Anthropic's development assistant, reaches a new level of autonomy. Cat Wu, former product lead of Claude Code, announced on November 3, 2026, that she left Anthropic after two years spent evolving the tool from internal use to a more autonomous development software. In her message, she thanks the teams of Claude Code, Cowork, and Tag, and highlights that user feedback shaped much of the deliveries.
Wu joined Anthropic in August 2024 as a product manager on Research PM. She recounted in March 2026 having used Claude Code internally from the start to build Streamlit applications, launch evaluations, and create reinforcement learning environments, without writing code by hand. According to her testimony, successive model versions also allowed going further: Opus 4 sometimes managed to add a table to Excalidraw in June 2025, then Opus 4.6 made this type of request reliable enough for public demonstrations.
When autonomy becomes the default setting
Anthropic formalized this shift with its auto mode, announced on August 7, 2026, then updated on October 22, 2026. On the Pro, Max, and Team plans, new sessions can run in auto mode, with an action ranking designed to block those deemed irreversible, destructive, or outside the user's scope. Actions judged safe pass, others trigger an approval request or a more cautious detour.
The system does not eliminate risk. Anthropic specifies that classification does not replace human verification, especially for sensitive production changes. Three consecutive blocks, or twenty in one session, revert Claude Code to manual approval mode. This mechanism changes the product balance: fewer interruptions for routine tasks, but always a human brake for operations that could break something valuable and live.
What this says about the agent market
The trend is clear: coding agents are no longer evaluated only on their generation quality but on how they fit into a company's infrastructure. Banks want to limit credential exposure and reduce attack surfaces. Publishers seek to make autonomy acceptable without turning every action into operational Russian roulette. The battle is therefore as much about security as deployment architecture.
For companies, the message is quite clear. A useful but poorly isolated agent can quickly become a governance problem, not just a productivity one. And the tools that will succeed long-term will likely be those that can do both: accelerate teams without asking them to leave the server door ajar.
Key points
- Goldman Sachs limits Claude Code for bankers and traders.
- Bloomberg cites an exception for certain "strats" profiles.
- Replit warned on November 3, 2026, about employee endpoints.
- Anthropic activated auto mode on Pro, Max, and Team.
- Cat Wu left Anthropic on November 3, 2026.
- Three blocks or twenty blocks revert Claude Code to manual mode.
By the numbers
- November 3, 2026 — date of Cat Wu's message and Replit's alerts.
- August 2024 — Cat Wu's arrival at Anthropic, according to her public account.
- August 7, 2026 — initial announcement of Claude Code's auto mode.
- October 22, 2026 — update of the auto mode deployment.
- 20 blocks — threshold reverting Claude Code to manual mode in a session.