Software company Figma's engineering team has built AI agents that help security engineers investigate alerts, search past incidents, check company systems, and prepare code fixes, according to a recently published case study on InfoQ. The agents draw lessons from earlier investigations, cutting down on repetitive tasks and helping engineers handle complicated alerts roughly 70% faster, though human review and tight controls stay in place. The system shows how automation can speed up security work without removing people from critical decisions.
The security framework runs on Panther SIEM and investigates alerts while examining audit logs from AWS, Okta, GitHub, GCP, and osquery, an open-source tool that queries computers for security and system data using SQL. It pulls information from more than 100 additional sources and can open pull requests. The team reports that complex alert resolution time dropped by about 70%, and on-call pages fell by 20% after the system lowered the severity of certain alerts. In a companion article about vulnerability detection, the engineering team noted that its agents uncovered more than 100 previously unknown vulnerabilities, including two critical flaws that traditional tools missed, and that its code reviewer hit 80% precision within a month. The system also boosted detection of known bugs by roughly 30% through a second review step, and the team saw about a 50% reduction in some coding errors after adding automated guidance.
Matthew Sullivan, formerly a security engineer at Figma and now at Nition, and Brad Girardeau, security engineering manager at Figma, write that the alert triage agent, which uses a model like Claude Opus, is where most of the investigation takes place. It receives the complete Slack thread history as context, its own steering memory, and a set of tools limited to what a security on-call engineer typically needs during triage. The article explains that the agentic system relies on AWS Bedrock Knowledge Bases, Amazon Kendra, Tines, and a Snowflake-based tool to search historical alerts and investigate Panther data. Sullivan and Girardeau add: "Memory ended up being the thing that had the most impact on how useful the system became over time." Three kinds of memory help the system improve investigations: past alerts, behavioral guidance, and learned database structures. Safety controls are embedded in the tools themselves, with agent-created pull requests set to draft by default and prompts written to stop sensitive data from being shared in public Slack channels.
The authors stress that the specifics depend on company size, the risks faced, and the feedback loops already running, warning they can't prescribe an exact approach. They do offer one central lesson: improve precision before recall. The order feels counterintuitive, they note, because the historical bugs already on hand can only measure recall and barely help with the precision that must be fixed first. As security teams hand AI agents more responsibility, the question of human approval remains open. Cloud security company Wiz recently reported in "GhostApproval: A Trust Boundary Gap in AI Coding Assistants" that six AI coding assistants could be tricked by malicious repositories while displaying what looked like a harmless approval prompt to users. InfoQ also recently covered OpenAI's disclosure about sandbox escapes. Figma acknowledges that the existing AI agents aren't perfect, but neither are humans, and concludes that the choice isn't between the two, with the right balance between automation and human oversight still evolving. Organizations adopting similar systems will need to decide where automation delivers value and where human judgment remains essential, a trade-off that will likely shift as both capabilities and risks mature.