- Published on
How Figma Uses AI Agents for Security
- Authors
- Name
- DeepHorus
Software company Figma has implemented AI-powered agents to significantly boost its security operations, reducing complex alert resolution time by approximately 70% and cutting on-call pages by 20%. The engineering team documented their approach in a recent blog post, detailing a system built on Panther SIEM that automates investigations across multiple platforms, including AWS, Okta, GitHub, GCP, and osquery. This move marks a major step in using agentic AI to handle repetitive security tasks, allowing human engineers to focus on more complex threats.
The core of the system is an alert triage agent, powered by a model like Claude Opus, which receives full Slack thread context and has access to a suite of tools. These tools enable it to search historical alerts, investigate data, query over 100 sources, and even open pull requests for code fixes. The integration uses AWS Bedrock Knowledge Bases, Amazon Kendra, and Tines, creating a robust investigative framework that operates within strict safety controls.
A key innovation is the agent's use of three distinct types of memory: past alerts, behavioral guidance, and learned database structures. The authors, Matthew Sullivan and Brad Girardeau, emphasized that separating these memory types was crucial for improving the system's usefulness over time. Safety is maintained by defaulting agent-created PRs to draft status and designing prompts to prevent sensitive data exposure.
In a related effort, Figma's agents have uncovered over 100 previously unknown vulnerabilities, including two critical flaws missed by traditional tools. The code reviewer agent achieved 80% precision within a month and improved detection of known bugs by about 30%. The team also reported a 50% reduction in certain coding errors after adding automated guidance, highlighting the tangible security benefits of this approach.
The company advises that while exact implementations vary, a primary lesson is to improve precision before recall. This counterintuitive order is because historical bugs mainly measure recall, whereas precision must be fixed first for effectiveness. Figma acknowledges its AI agents are not flawless, but argues that human processes are also imperfect, framing the challenge as finding the right balance between automation and oversight.
As the industry adopts similar technologies, questions about trust and safety remain paramount. Recent reports from Wiz and OpenAI about vulnerabilities in AI coding assistants underscore the risks of over-reliance. Figma concludes that the evolving choice is not between AI and humans, but about integrating them thoughtfully to enhance security without compromising control.
