TL;DR
- Goldman Sachs revoked access to Anthropic's Claude Code and Microsoft's GitHub Copilot for traders and investment bankers in October, restricting both tools to engineers only. Quantitative strategists ("strats") retain access. Goldman cited "productivity and risk management" without specifics. Traders had used the tools for months to monitor client trading flows, model market positions, and produce content. Goldman's CIO had praised "vibe coding" by non-technical staff in July 2026. An internal Goldman partner warned in August that AI coding dependence could degrade bankers' reasoning skills. The October rollback reversed months of broader deployment. - Goldman maintains its own internal AI assistant (combining OpenAI, Anthropic, and Google models) and has not restricted that. The action applied specifically to Claude Code and GitHub Copilot — the external coding tools with the highest code-generation capability. Earlier in 2026, Goldman restricted Claude access for Hong Kong staff over concerns about confidential data leaving the firm through third-party AI services. - A WIRED investigation published November 2 documented a network of more than 400 apps and domains that posed as ordinary AI services — pet portrait generators, language tutors, workout coaches — while hiding tools to generate sexually abusive deepfakes. The apps sent device and referral data to external servers, which determined whether to show the harmless version or load the explicit interface. 350,000+ downloads. Many rated suitable for all ages. Promoted through thousands of explicit ads on Meta's platforms. Traced to a Hong Kong entity named KeLing Technology Limited. Apple removed 19 apps after the investigation published. Meta took action against 11,600 ad accounts and blocked 209 apps from advertising. - The deepfake network used a server-side trigger to evade app store review: the app itself contained no explicit content; the harmful interface loaded remotely based on user data. Standard app store review, which examines the code and content in the submitted binary, cannot detect server-side features that activate conditionally.
Two failures of AI governance this week — one inside a Wall Street bank, one in the consumer app ecosystem. In both cases, the capability was deployed before the oversight was in place. In both cases, the oversight arrived only after the harm potential was visible.
Goldman Sachs: from "vibe coding" to engineers-only in four months
The timeline is worth reconstructing precisely.
July 2026: Goldman CIO Marco Argenti told a public interviewer that non-technical staff — traders, bankers — were engaging in "vibe coding," building rough prototypes with AI tools and then handing them to engineers for refinement. He described it approvingly: traders having an idea, building a quick prototype with AI, then handing it off. This was the firm's stated strategy.
August 2026: Goldman partner Chris Churchman, leading one of the bank's flagship AI projects, warned publicly that the technology could degrade bankers' reasoning skills if relied on too heavily. This is the first signal that the internal view was not uniform.
October 2026: Internal guidance updated. Traders and investment bankers lose access to Claude Code and GitHub Copilot. Engineers and strats retain access. The access change applied to the two tools specifically designed to write, edit, and generate production-quality code.
The Goldman spokesperson's characterization: "After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term."
What this means in plain terms: the productivity gains from giving coding tools to non-coders did not outweigh the risk management concerns that emerged from the experiment. The experiment ran for months. The rollback happened in October.
What the risk management concerns likely were, based on available information: FINRA requires broker-dealers to maintain supervision, recordkeeping, and cybersecurity controls regardless of what tools employees use. A trader who writes code using Claude Code and deploys a model that monitors client trading flows has created a piece of financial infrastructure. That infrastructure is subject to the same regulatory requirements as infrastructure built by engineers. The question of whether traders' Claude Code-generated models met those requirements — supervision, documentation, cybersecurity review — is almost certainly the risk management issue that Goldman is not disclosing.
Earlier in 2026, Goldman restricted Claude access for Hong Kong staff over concerns about confidential information leaving the firm through third-party AI services. The same concern applies at a larger scale to any firm employee using Claude Code or GitHub Copilot: the model is being served by Anthropic or Microsoft infrastructure, and prompts and code may be logged or used by those providers depending on the terms of service and configuration. Goldman's enterprise agreements likely address this, but the compliance and audit requirements in a regulated financial institution are considerably more demanding than in a typical software company.
The practical effect of the rollback: Goldman's roughly 12,000 engineers retain the tools. Strats retain access. The people who were using the tools to do non-engineering work — traders using code generation to accelerate trading flow monitoring — are back to either delegating to engineers or using Goldman's internal AI assistant, which runs on the same underlying models but through Goldman's own infrastructure with Goldman's own data controls.
The enterprise lesson: "vibe coding" as a strategy assumes that the output of non-engineers using code generation tools has the same quality and compliance profile as engineer-built code. Goldman's experiment found that the risk management surface of non-engineer-generated code in a regulated environment is different. The access is being tiered accordingly.
The counterpoint: this is one bank's decision about one set of tools. GitHub Copilot is in public preview for desktop computer use. OpenAI's Codex and Astra are being used by developers at millions of companies. The Goldman rollback is a data point about financial industry risk tolerance, not a verdict on whether AI coding tools work.
The 400+ deepfake app network: what server-side evasion looks like at scale
The WIRED investigation (published November 2, conducted by WIRED and AI Forensics research group) documents a different kind of governance failure — one where the standard review mechanisms were structurally incapable of catching what was happening.
The network: more than 400 apps and domains, connected through shared infrastructure and a JavaScript bundle containing links, encryption keys, and ad trackers. The apps presented as ordinary AI services: pet portrait generators, language tutors, fitness coaches. They were listed on Apple's App Store and Google's Play Store. They accepted payments through Apple Pay and Google Pay. They were advertised through explicit ads on Meta's Facebook and Instagram. Many were rated suitable for all ages. Together they had more than 350,000 downloads.
The evasion mechanism: after installation, each app contacted an external server and sent data — the user's language preference, location, device identifier, and how they found the app (their referral source, including which ad they clicked). The server determined, based on that data, whether to show the advertised harmless service or to load a separate web interface containing the explicit deepfake generation tools.
The web interface, when loaded, offered dozens of templates for generating sexually explicit images and videos from real photographs users upload. Templates included depictions of sexual violence. WIRED described template labels including "bedroom rape" and "threatening knife."
Why standard app review could not catch this: app store review examines the binary submitted to the store — the code, the content, the permissions requested. The deepfake interface was not in the binary. It loaded remotely, triggered by server-side logic that checked user characteristics before deciding what to show. The app as submitted to Apple and Google contained nothing that would fail a content review. The harmful functionality existed only in what the server returned after determining the user was a valid target.
This is a documented technique. Researchers who study app store evasion have called it "server-side feature flagging" used maliciously. Legitimate apps use server-side feature flags to roll out new features gradually. The same mechanism, used with malicious intent, creates apps that behave differently in review than they behave in deployment. Apple's reviewers see the harmless version. Users who found the app through an explicit ad on Meta, or who have characteristics the server identifies as relevant, see the deepfake interface.
The developer trail: WIRED traced the network to a Hong Kong-registered entity named KeLing Technology Limited and Chinese-language developers. The entities scrubbed their online presence after being contacted by WIRED.
Platform responses after disclosure: Apple removed 19 apps. Meta took action against 11,600 ad accounts, 4,700+ business pages, and 15,000 pages, and blocked 209 apps from advertising. These numbers suggest the advertising infrastructure for the network was more extensive than the app network itself — the ads were the primary discovery mechanism.
The structural problem: the apps started appearing in late 2025. New domain registrations peaked at approximately 30 per week during spring 2026. The network operated for roughly a year before the investigation. The evasion technique that made this possible — conditional server-side feature loading based on user characteristics — is not new and not difficult to implement. The app stores' review processes, which examine submitted binaries, were not adapted for this attack surface.
What both stories say about the same month
Goldman's rollback and the deepfake network investigation are not the same kind of event. One is a large regulated institution making a measured governance decision after a controlled experiment. The other is a criminal network operating on a global scale using deceptive infrastructure to distribute tools for generating sexually abusive imagery.
They are addressed in the same week because they both illustrate the same underlying gap: the frameworks for governing AI deployment — enterprise policy, app store review, financial regulation — were built for a different kind of technology risk.
FINRA's broker-dealer supervision rules were not written for an environment where a trader can generate production-quality trading infrastructure in an afternoon using an AI model. App store review was not designed for apps that contain no harmful content in the submitted binary but load it remotely after deployment. Goldman's experimentation period discovered the former; WIRED's investigation revealed the latter.
Both problems are solvable with the right controls, and both are actively being addressed — Goldman with its tiered access model, the app stores with belated removals. The question is whether the controls are keeping pace with the capability. The Goldman experiment lasted four months before the rollback. The deepfake network ran for a year. In both cases, the capability was deployed before the oversight was ready. In both cases, harm was limited only by the speed of the discovery.
