Anthropic Warns AI Industry Has No Brake Pedal as Claude Codes 80% of Its Own Codebase

Anthropic co‑founder Jack Clark has a metaphor for where the AI industry stands right now. Speaking to,¹ he said: "When I look down at the car we're driving, all I have is a gas pedal. I don't have a brake pedal."

Clark laid out a stark warning in the June 5 interview — the industry is hurtling toward a future where AI systems can improve themselves without human oversight, and there's no mechanism to slow things down when needed.

The timing is notable. Anthropic recently filed for an IPO and is actively building massive compute infrastructure, including a $35 billion chip financing deal. This is not a company asking to slow down because it's falling behind — it's the opposite. The warning comes from one of the best‑positioned labs, which makes it harder to dismiss.

The numbers tell a dramatic story. According to MediaNama, which analyzed Anthropic's institute blog post, Claude now authors over 80% of code merged into the company's production codebase — up from low single digits before Claude Code launched in 2025.

The typical Anthropic engineer merged 8× more code per day in Q2 2026 compared to 2024. Human engineers have shifted from writing code to reviewing AI outputs and choosing which problems to tackle. As the blog post puts it: "The doing — writing the code, running the experiment, producing the result — now costs almost nothing in human time."

Jack Clark told ³ that Claude could handle 100% of Anthropic's coding tasks within a couple of years. The trajectory is steep and accelerating.

The core concern in Anthropic's warning is recursive self‑improvement — AI systems that can design and build their own successors without human direction. In a blog post co‑authored by Clark and Marina Favaro of The Anthropic Institute, they describe a near future where AI models independently run experiments, propose hypotheses, and conduct open‑ended research.

Current Claude models have already progressed from tasks that took humans minutes in 2024 to tasks lasting up to 12 hours, per.² The concern isn't that today's systems are about to escape — it's that the progression curve points toward capabilities arriving faster than governance mechanisms can keep up.

Clark compared the situation to Cold War arms control, told:¹ "We've done this before. In the height of the Cold War, under highly tense situations between rivalrous countries, they found ways to stabilize aspects of the nuclear arms race."

Anthropic is not asking for a unilateral slowdown — it's calling for a coordinated mechanism with verification capabilities. The company stated it would slow down or temporarily pause "if other developers at or near the frontier also did so in a verifiable manner," as reported by MediaNama.

Clark went further with,³ specifying he is not advocating voluntary slowdowns but urging the creation of actual regulatory mechanisms with enforcement power.

Anthropic outlines three trajectories: deceleration (natural slowdown), automated development with human direction (concerning), and full recursive self‑improvement (highest concern). According to MediaNama, the lab is most worried about scenarios two and three, which could give governments and societies "little time to adapt."

Not everyone is buying the warning at face value. Abeba Birhane, an AI accountability expert at Trinity College Dublin, dismissed the call as a "clever marketing trick" in comments to the.⁴ Her argument: current AI systems are not autonomous and require constant human verification.

"They are presenting their AI systems as if they are fully autonomous, but this is incorrect," Birhane told the Irish Examiner. "Their top models continually need human verification for fine tuning. They are not reliable systems. There are layers of domain experts and contractors that verify code."

She accused Anthropic of using the warning to "abdicate responsibility to their own neural network" and "evade accountability for the people designing the system." Ian Dodson, CEO of AICertified, offered a different read — he sees it as a political repositioning after the Trump administration cut off federal use of Claude. "They've gotten their ass kicked by the US government, so now they're in the anti‑Trump camp," Dodson told the.⁴

Anthropic's safety positioning has put it in direct conflict with the US government. In February 2026, President Trump instructed federal agencies to stop using Claude after Anthropic refused to remove safety precautions that would allow autonomous lethal decision‑making, per the.⁴ OpenAI's ChatGPT replaced Claude in those agencies.

Yet the relationship is more complicated than a clean break. As Gizmodo reported this week, the NSA is reportedly using Anthropic's Mythos AI model for offensive cyber operations — despite Defense Secretary Pete Hegseth's opposition. The White House Chief of Staff reportedly authorized the NSA to continue using Mythos.

This tension — between Anthropic's safety‑first public stance and the government's operational use of its models — forms the backdrop for the brake pedal argument. If even a safety‑focused company can't control how its models are deployed, what happens when systems become truly autonomous?

The immediate takeaway for developers is practical: Claude is already doing the majority of coding at one of the world's top AI labs. The same tools — Claude Code, and similar systems — are available to builders today. The productivity gains Clark describes (8× more code per engineer) aren't theoretical.

But the longer‑term signal matters too. If Anthropic succeeds in establishing a brake pedal mechanism, it could mean regulatory frameworks that affect which models are available, under what conditions, and with what safety restrictions. Builders who depend on frontier models should pay attention to these governance debates — they'll shape what you can deploy to production.

The debate also highlights a real split in the AI community. Some, like Birhane, argue the autonomy is overstated — the systems still need humans in the loop. Others, like Clark, see the trajectory and want guardrails before the industry crosses a point of no return. Both perspectives have implications for how builders should think about AI dependency in their own stacks.