Figma's AI Agents Accelerate Security Investigations by 70%

Figma's security engineering team has released details on their approach to using AI agents for investigating security alerts, searching incident history, auditing systems, and generating code fixes. The system cuts resolution time for complex alerts by roughly 70% and trims on-call pages by 20%, leveraging machine learning to handle repetitive work while preserving human oversight throughout the process.

Building on Real-World Alerts#

The foundation is a security infrastructure centered around Panther SIEM that monitors and investigates alerts across multiple platforms. The agents query AWS, Okta, GitHub, GCP, and osquery (an open-source SQL-based tool for security and system diagnostics) alongside more than 100 additional data sources. The system can even create pull requests for potential fixes.

At the core sits an alert triage agent powered by models like Claude Opus. As the engineering team explains, this agent receives the full history of a Slack conversation about the alert, its own accumulated memory, and a curated set of tools typically available to an on-call security engineer. The architecture draws on AWS services including Bedrock Knowledge Bases and Amazon Kendra, plus Tines and a Snowflake-based data layer to search past alerts and dig into Panther's records.

Memory as Competitive Advantage#

A major breakthrough came from how the system learns from experience. According to Matthew Sullivan and Brad Girardeau of Figma's security team, "Memory ended up being the thing that had the most impact on how useful the system became over time." The agents maintain three distinct memory types:

• Past alerts: Historical context from previous investigations
• Behavioral guidance: Rules and patterns learned from how human engineers typically approach problems
• Database structures: Understanding of system schemas that improves query accuracy

Keeping these separate proved essential to the system's effectiveness and stability.

Safety Built Into the System#

Figma embedded security controls directly into the agent's tooling rather than relying solely on supervision. Agent-generated pull requests default to draft status, preventing accidental merges. Prompts are designed to block sensitive data from being broadcast in public Slack channels. This layered approach lets the system operate efficiently while staying within safety guardrails.