{
  "id": 5932446,
  "title": "How Figma Uses AI Agents for Security",
  "url": "https://urgent.news/2026/09/06/how-figma-uses-ai-agents-for-security",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-06T06:59:00.000Z",
  "source": {
    "name": "InfoQ",
    "slug": "infoq",
    "url": "https://www.infoq.com/news/2026/09/figma-security-agents/"
  },
  "original_language": "en",
  "account": "Software company Figma recently shared how their engineering team constructed AI agents to bolster the security team's ability to investigate alerts, sift through past incidents, probe company systems, and even draft code fixes. These agents, powered by Panther SIEM, learn from prior investigations, thereby cutting redundant tasks and accelerating resolution of intricate alerts by nearly 70%, while retaining human oversight and robust safeguards.\n\nThe security investigation system, built atop Panther SIEM, probes alerts and audits logs across AWS, Okta, GitHub, GCP, and osquery, a SQL-enabled open-source tool that digs up security and system information from over 100 sources, even opening pull requests. The authors detail that the alert triage agent, employing a model akin to Claude Opus, shoulders most of the investigation process. It receives the entire Slack thread history as context, its own memory steering mechanism, and a bespoke toolkit tailored to the typical on-call engineer's needs during triage.\n\nThe system leverages AWS Bedrock Knowledge Bases, Amazon Kendra, Tines, and a Snowflake-based instrument to scour historical alerts and scrutinize Panther data. The authors highlight that memory significantly amplified the system's utility over time. They identified three types of memory: historical alerts, behavioral guidance, and learned database structures, each enhancing the system's capacity to refine investigations progressively. Critical safety protocols are embedded within the tools, with agent-generated PRs defaulting to draft status, and prompts meticulously crafted to ward off the inadvertent disclosure of sensitive data in public Slack channels.\n\nIn another piece, \"How Figma stays ahead of vulnerabilities with agents,\" the engineering team disclosed that their agents unearthed more than 100 previously unknown vulnerabilities, including two critical flaws that evaded traditional detection methods. Notably, their code reviewer achieved 80% precision within a month, and the system boosted known bug detection by roughly 30% through a secondary review stage. These enhancements resulted in about a 50% reduction in specific coding errors after integrating automated guidance.\n\nWhile the authors outline the steps they took to attain these enhancements, they caution that there is no one-size-fits-all solution: \"We can't tell you exactly what to do: The specifics hinge on your company's scale, the risks you confront, and the feedback loops you already employ.\" Nevertheless, they underscore a pivotal lesson: prioritize precision over recall. This counterintuitive advice stems from the observation that historical bugs, while crucial for measuring recall, offer limited aid in achieving the precision required to address emerging issues effectively. As security teams increasingly entrust AI agents with greater responsibilities, the role of human approval remains a fluid and open question. Cloud security firm Wiz recently warned in \"GhostApproval: A Trust Boundary Gap in AI Coding Assistants\" that six AI coding assistants could be duped by malicious repositories, all while presenting users with innocuous approval prompts. Similarly, OpenAI recently disclosed instances of sandbox escapes. Figma acknowledges the imperfections of their current AI agents, echoing the broader acknowledgment of human fallibility. However, they emphasize that the quest for the optimal equilibrium between automation and human oversight remains an evolving endeavor.",
  "summary": "The engineering team at software company Figma recently documented how they built AI agents to help their security team investigate alerts, search past incidents, check company systems, and even prepare code fixes. The agents learn from previous investigations, reducing repetitive work and helping engineers resolve complex alerts about 70% faster. By Renato Losio",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}