Anthropic's AI Agents Were Merging 30,000 Times a Day, So They Replaced Git

by Sevda Polat — October 14, 2026

Anthropic just revealed it ripped out Git entirely from its main codebase and replaced it with a custom version control system called Scion, built by seven people in five months. The reason: Claude agents and automated jobs now account for almost all users of the monorepo, pushing daily merges from ~2,500 in May to 30,000+ in September. At that scale, with 100 GB of history and 25 million checkouts a day, naive Git traffic would have hit 600 petabytes daily. Scion cut infrastructure footprint by 200x. And in a nicely recursive detail, an internal Claude model rewrote Scion's own ingestion pipeline to run 4x faster.

Anthropic's AI Agents Were Merging 30,000 Times a Day, So They Replaced Git

by Sevda Polat — October 14, 2026

Anthropic just revealed it ripped out Git entirely from its main codebase and replaced it with a custom version control system called Scion, built by seven people in five months. The reason: Claude agents and automated jobs now account for almost all users of the monorepo, pushing daily merges from ~2,500 in May to 30,000+ in September. At that scale, with 100 GB of history and 25 million checkouts a day, naive Git traffic would have hit 600 petabytes daily. Scion cut infrastructure footprint by 200x. And in a nicely recursive detail, an internal Claude model rewrote Scion's own ingestion pipeline to run 4x faster.

AI Agents Are Breaking Things and Getting Paid

Columbus Day weekend was quiet for most of tech. Less so for anyone working in security.

Google raised Chrome's top bug bounty to $250K in April because full-chain browser exploits had become extraordinarily rare. The researchers worldwide who could chain one together might fill a single conference room.
Offensive security carries an inherent asymmetry: attackers need one way in, defenders need to cover every surface. That asymmetry has always been expensive to exploit — until the cost of capable reasoning dropped by orders of magnitude.
A top vulnerability researcher costs well into six figures a year. An AI agent's API bill for a week of comparable work? About $40K.

The economics of finding serious vulnerabilities are shifting permanently. The people who price bug bounties have been adjusting accordingly.

TODAY'S BIG MOMENT

AI Agents Are Breaking Things and Getting Paid

Columbus Day weekend was quiet for most of tech. Less so for anyone working in security.

Google raised Chrome's top bug bounty to $250K in April because full-chain browser exploits had become extraordinarily rare. The researchers worldwide who could chain one together might fill a single conference room.
Offensive security carries an inherent asymmetry: attackers need one way in, defenders need to cover every surface. That asymmetry has always been expensive to exploit — until the cost of capable reasoning dropped by orders of magnitude.
A top vulnerability researcher costs well into six figures a year. An AI agent's API bill for a week of comparable work? About $40K.

The economics of finding serious vulnerabilities are shifting permanently. The people who price bug bounties have been adjusting accordingly.

AI Chains Three Chrome Zero-Days, Earns $250K

XBOW researchers directed GPT-5.6 Sol agents to chain three Chrome zero-days — V8 sandbox escape, renderer escape, and OS-level code execution on Windows. Six days, ~$40K in API costs, $250K bounty. Humans made the key architectural calls. Roughly a billion Chrome users were in scope.

Linux Root and OTP Bypass Both Found by AI

Two more data points in a growing pattern: an AI-discovered Linux kernel flaw grants local root access and can escape a guest VM to the host. Separately, an AI agent found a critical OTP bypass in production. The writeup notes the exploit itself was the straightforward part.

LiteLLM Bugs Could Expose Your LLM API Keys

Vulnerabilities in LiteLLM — the popular unified gateway for routing calls across LLM providers — could let attackers steal your OpenAI and Anthropic API keys. If you're running a multi-model setup through LiteLLM, worth checking your deployment immediately.

Docker Hub Flaw Lets Attackers Impersonate Any User

CVE-2026-105445 allows attackers to impersonate Docker Hub users and tamper with container images. Given how many CI/CD pipelines pull images from Docker Hub without additional verification, the downstream exposure could be significant. Patch and audit your image sources.

Meanwhile in the Rest of the Tech World

Agentic Commerce

AI Agents Quietly Made 138 Million Crypto Payments