Goldman Sachs Pulls Claude Code From Traders After Four Months of Ungoverned AI Scripts
When a 12,000-engineer institution yanks AI coding permissions from traders after a single summer, every company running on loose AI-generated code is on notice: the absence of review, testing, and audit turns productivity into liability. The cheap engineering gates described here cost one afternoon to install, and they're the difference between keeping AI tools and losing them to a blanket ban.
In July, Goldman Sachs CIO Marco Argenti encouraged non-engineers to vibe code: let AI turn ideas into working tools, then hand them to engineers. By October, the firm had silently revoked traders' access to Claude Code and GitHub's AI coding tools. Bloomberg reported a revised internal policy restricting those tools to engineers after a productivity and risk assessment. The difference between a summer of encouragement and a fall of revocation came down to a missing layer: the code had no review, no tests, no audit, no rollback, yet it ran daily on machines handling client trade data.
That's an engineering problem, not an AI trust problem. A trust chain built from code review, automated tests, static checks, staged rollout, and monitoring/rollback treats any code the same regardless of author. AI code that passes those gates is as safe as human code; human code that skips them is as dangerous as a trader's script. Goldman's internal AI platform has governance built in, but traders were using external tools outside its radius.
The practical path for small teams: put all executable code in version control, install pre-commit hooks with secret scanning (gitleaks, detect-private-key), add semgrep for static analysis, and write down answers to three questions: who uses AI, with what data, under what terms. The comparison table in the piece covers six dimensions from code review to incident response, and the last row, traceability after an incident, is the one that often gets overlooked.
The same model that wrote the trader's script hasn't changed; what changed is that the code now runs inside a process with review and tests. That framing makes AI quality a non-issue and governance the only lever firms control.
Static analysis catches patterns that human review and AI alike gloss over: disabled certificate checks, permissive file defaults, shell injection. Because models learn from examples that include these mistakes, rules become a proxy for senior judgment.
Legal enterprise contracts and internal AI platforms both fail when employees prefer newer external tools; the time gap between governance updates and model releases is where risk accumulates.
Goldman's role as Anthropic's IPO underwriter while restricting Claude Code shows that security policy and business interests can run on separate tracks without contradiction.