Figma puts AI agents on security alert triage

Figma's security team now runs two AI agents in production: one takes the first look at incoming alerts, the other reads code. The company reports roughly 70 percent faster resolution on complex alerts and 20 percent fewer on-call pages. InfoQ summarized the account on September 6, 2026.

What each agent does

The triage agent inherits the Slack thread behind an alert plus the tool access an on-call engineer would normally reach for: queries against the Panther SIEM, audit logs, and checks on internal systems. Figma names Claude Opus as the model behind it.

The second agent reviews code and can prepare fixes. Within a month it reached 80 percent precision and turned up more than 100 previously unknown vulnerabilities, including two critical flaws that conventional tooling had missed. A second review pass improved detection of already-known bug classes by about 30 percent, and automated guidance pushed into the developer workflow cut certain error types by about 50 percent.

Memory mattered more than the model

The team's own framing is that the biggest lever was not model choice but memory. The system retains three kinds of knowledge: past alerts with their investigation trail, behavioral guidance on how to run an investigation, and learned structures of the databases it queries. That accumulated context, the engineers say, is what made the system more useful over time.

Behind the agents sit more than 100 data sources, among them AWS, Okta, GitHub, GCP and osquery, wired together with components such as AWS Bedrock Knowledge Bases, Amazon Kendra, Tines and Snowflake-based tooling. The account draws on Matthew Sullivan, formerly a security engineer at Figma and now at Nition, and Brad Girardeau, a security engineering manager at Figma.

Where the human stays

Pull requests written by an agent open as drafts, not as finished proposals. Prompt-level rules are meant to keep sensitive data out of public Slack channels. Approval is explicitly a human step.

The stated design rule is to improve precision before recall: an eager agent that flags everything simply moves the workload rather than reducing it. The write-up also flags a risk beyond Figma — AI coding assistants can be led astray by deliberately malicious repositories.

What is not established

Every figure here comes from Figma's own account and has not been independently verified. Absolute alert volumes, the baseline period the improvements are measured against, and the running cost of the system are not disclosed.

Only one independent newsroom covered this for our purposes. How many false positives triage produces, and how often a human overrides the agent's read of an alert, is not stated in the source.