Goldman Sachs limits Claude Code: security lessons for SMEs

Goldman Sachs limits the use of Claude Code: what's behind it?

Recently, The New Stack reported on a news story that has sparked debate in the financial sector: Goldman Sachs has restricted access to AI coding tools such as Anthropic's Claude Code and Microsoft's GitHub for certain employees, specifically bankers and traders, while strats (quantitative employees) retain access. This decision, reported by Bloomberg, highlights growing concerns about the security of AI agents operating on employees' devices.

Amjad Masad, CEO of Replit, has pointed out on social media that AI agents on employee endpoints are becoming a systemic risk for companies and governments. According to Masad, this is a widespread trend in the financial sector: "the ban hammer is falling." Although no specific incidents have been detailed, the concern centers on credential exposure, malicious instructions, and compromised dependencies.

The problem of credential exposure

Running coding agents directly on employees' machines raises serious security questions. Agents can read and reuse long-lived credentials, making supply chain attacks easier. In fact, Replit has published a security advisory recommending keeping AI agents off employees' devices and running them in company-controlled environments, with short-lived credentials and audit logs.

The advisory mentions recent incidents, such as a compromised Claude Code extension distributed through the Cursor marketplace, and an attack that hijacked approximately 1,900 GitHub accounts before targeting some 30,000 repositories. These examples underscore that a single weak link can have a much greater impact than it seems.

Is restricting access to developers enough?

Limiting AI tools to engineers only does not necessarily solve the problem. Even experienced developers can run agents that follow malicious instructions or incorporate compromised software. Research has shown how easy it is to divert an agent: one team found that removing a single line from Claude Code's instructions increased its out-of-scope actions.

That is why Replit argues that companies must control what agents can do, not just who can use them. This involves isolated environments, separate credential management, audit logs, and approvals.

Implications for SMEs and IT teams

Goldman Sachs' decision may seem remote for an SME, but the lessons are applicable. Many small and medium-sized businesses are adopting AI tools to increase productivity, often without adequate safeguards. Using coding agents on employees' laptops can expose access credentials to critical systems, such as email, CRM, or databases.

Moreover, the promise that non-developer employees can create software with these tools is attractive, but it becomes risky when agents have access to sensitive corporate systems. Security must be a priority.

Recommendations for safe adoption

At ForgeNEX we believe that SMEs can benefit from AI without compromising security. Here are some practical recommendations:

• Run agents in controlled environments: Whenever possible, use isolated cloud environments managed by the company, instead of running agents directly on employees' devices.
• Limit privileges: Apply the principle of least privilege. Agents should have access only to the resources necessary for each task.
• Use short-lived credentials: Avoid long-lived credentials that can be read and reused by agents. Implement ephemeral tokens and periodic rotation.
• Audit and monitor: Keep audit logs of agents' actions and periodically review their behavior to detect anomalies.
• Train employees: Awareness of security risks and good practices when using AI tools.
• Review generated code: As we discussed in our article on code review with AI, it is essential to verify what agents produce and maintain human oversight.

Security should not be a brake on innovation, but it should be a priority. Goldman Sachs' experience reminds us that even organizations with large resources are reevaluating the balance between productivity and risk. SMEs must learn from these lessons and adopt a proactive approach.

Conclusion

The restriction of Claude Code at Goldman Sachs is a wake-up call. AI agents are powerful tools, but their indiscriminate use on employees' devices can open security gaps. SMEs must implement adequate controls, such as isolated environments, credential management, and auditing, to take advantage of their benefits without exposing themselves to unnecessary risks.

At ForgeNEX we help companies design and implement secure AI solutions tailored to their needs. If you are considering adopting coding agents, do not hesitate to contact us to evaluate your case.

Source: The New Stack. Analysis and adaptation: ForgeNEX.