Figma Uses Security Agents to Speed Complex Alert Response, Keeps Fixes as Drafts

Figma has put AI agents into security investigations that reach from alerts and historical incidents to company systems and proposed code fixes. The company says the system cut resolution time for complex alerts by about 70%, but it has kept agent-created pull requests in draft mode and added protections against sensitive information reaching public Slack channels.

An investigation layer across the security stack

The system uses Panther SIEM to investigate alerts and check audit logs across AWS, Okta, GitHub, GCP and osquery. It can query more than 100 additional sources and open pull requests, tying security triage to both the records needed for an investigation and a potential remediation path.

Figma says its alert-triage agent receives Slack-thread history, steering memory and tools scoped to normal security on-call tasks. It describes the model as one like Claude Opus; AWS Bedrock Knowledge Bases, Amazon Kendra, Tines and a Snowflake-based tool support searches of historical alerts and Panther data.

What the agent retains between investigations

• Past alerts provide context from earlier cases.
• Behavioral guidance steers how the agent conducts an investigation.
• Learned database structures help it navigate the information it searches.

The authority boundary is built into the workflow

Figma has put constraints alongside the agent’s tools. Pull requests created by agents are drafts by default, rather than ready-to-merge changes, while prompts are designed to prevent sensitive investigation data from being shared in public Slack channels.

Broader security claims remain company-reported

In separate work, Figma reported that its agents found more than 100 previously unknown vulnerabilities, including two critical flaws missed by traditional tools. It also said its code reviewer reached 80% precision within a month, improved detection of known bugs by about 30%, and cut some coding errors by about 50% after automated guidance was added.

Those outcomes extend the case for using agents in security work, but they do not settle how much approval should remain with people. Figma acknowledges its agents are imperfect and says the appropriate balance between automation and human oversight is still evolving.