Three stories broke in the past 72 hours that, taken together, capture a shift in tone across the AI industry — from unchecked enthusiasm toward something more cautious, even defensive.

Goldman Sachs hits the brakes on "vibe coding"

After months of letting non-engineers loose on powerful AI coding assistants, Goldman Sachs has reversed course. Goldman Sachs is pulling AI coding tools like Claude Code and GitHub Copilot from traders and bankers, limiting access to engineers only, after CIO Marco Argenti had publicly praised "vibe coding" by non-technical staff as recently as July. The bank's official line points to "a variety of factors including productivity and risk management."

The about-face is notable because Goldman had been a poster child for front-office staff using AI to prototype tools before handing them to engineers. Non-engineers had been independently assembling functional solutions before handing them off to the firm's 12,000 professional developers for refinement. Now that pipeline is closed to everyone outside engineering, even as Goldman rolls out its own proprietary GS AI Assistant to roughly 46,000 employees while running ongoing trials of agentic AI, including in trade surveillance. It's less a retreat from AI than an attempt to bring it in-house and under tighter control.

Anthropic's own IPO filing warns about AI building AI

In a far more existential vein, Anthropic used an amended IPO prospectus — not a blog post — to flag a risk at the heart of its business. An amended S-1A dated November 2, 2026, puts Anthropic's concern about AI-assisted R&D into its formal risk disclosures, with forecasts suggesting the company may soon meet its policy definition of dramatically accelerated progress.

The concern centers on recursive self-improvement: Anthropic's Responsible Scaling Policy sets two triggers for dramatically accelerated AI progress — an AI system performing an entire AI research team's work at competitive cost, or automated AI research plausibly doubling aggregate capability progress against a baseline. Crucially, Anthropic has not designated any current model as exceeding the threshold and is still assessing the underlying data. Still, the company acknowledges the core tension: it is considering policy changes to keep key research processes under human supervision as capabilities may advance faster than safety evaluations, alignment research and safeguards. Putting this in a securities filing — rather than a research paper — signals Anthropic now treats the risk as material enough for investors to weigh.

Mistral bets on open weights at the trillion-parameter scale

On the more optimistic side of the ledger, French lab Mistral pushed further into open-source territory. French AI company Mistral has released the weights for its trillion-parameter open source AI model Mistral Large 4, allowing businesses and governments to download and deploy the system themselves. Nicknamed "Le Chonk," the roughly 1 trillion-parameter model uses 49 billion active parameters per token, reads text and images, and handles a 1 million-token context.

Mistral explicitly pitches the release around defense use cases: "Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally," the company said. Extra post-preview tuning paid off on benchmarks too — three more weeks of reinforcement learning took the model from 28% to 41% on Terminal-Bench 4.0, with the weights now on Hugging Face.

The common thread

Whether it's a bank clawing back developer tools, a frontier lab airing its own risks to regulators and investors, or an open-weights release aimed squarely at cybersecurity, the message is consistent: the AI industry is moving from "ship it everywhere" to "figure out who should actually be touching this."

Sources

AnthropicArtificial IntelligenceDaily AI UpdateGoldman SachsMistral AIOpen Weight Models