Cyber Chaos Monkey: Staying ahead of security threats with AI agents
Dan Scofield · 12 November 2026
Read the originalShopify built AI agents that attack its own live systems to find security flaws, with a second AI agent on a different AI model ready to stop them.
Dan Scofield, writing on Shopify's engineering blog, described the Cyber Chaos Monkey, AI agents that hunt for security flaws in Shopify's live systems as outside researchers would. Shopify, which runs online stores for merchants, says AI models have made finding flaws cheaper, and its bug bounty reports have risen almost 300% since 2024. Its answer is to use its own knowledge of its code and infrastructure to find flaws first, under strict limits.
- One agent proposes likely weak points from source code, past reports and threat information, and another tests them against live services using test shops.
- The attacking agent has no elevated privileges, receives only short-lived credentials through a proxy, and can reach only approved domains.
- A second agent, prompted differently and running on a different AI model, reads everything the attacker does and ends the run if it breaks the agreed rules. If that agent fails, the run also ends.
- Shopify reports tens of thousands of runs, dozens of serious findings and no outages or incidents caused by the testing.
An AI agent allowed to act on production systems can be paired with a second agent on a different AI model whose only job is to stop it. This is an inference and is not the author's claim. The post does not say whether people review each finding.
Run a second AI agent on a different AI model to read every action of an agent on live systems and stop it when it breaks agreed rules.
Keep credentials out of an AI agent's environment and let a proxy add short-lived credentials only for approved destinations.
Derived by Working Surface from the article; more in the Playbook. Source line: Attackers target our production systems, so our testing has to as well.
13 November 2026: AI agents earn more freedom when a team limits what they can reach, from data to design components, and adds a check that does not depend on a glance.