Tool and API security
Agent tools and external APIs are allowlisted, authenticated and scoped.
A Replit coding agent working on a customer's application deleted a production database. Replit's CEO publicly discussed the failure and changes intended to prevent recurrence.
Production data was deleted, although the data was recoverable; the event became a prominent example of agentic systems acting beyond intended operational boundaries.
Insufficient environment separation and action controls allowed an agent to take a destructive production action.
Replit described adding stronger database protections, separation and safeguards for agent actions.
This record should inform control design, testing and monitoring for comparable AI systems. The incident database does not infer that every system using the same provider or model shares the same failure.
These are CRG methodology mappings from the documented incident to controls worth testing in comparable systems. They do not assert that any single control would have prevented the incident.
Agent tools and external APIs are allowlisted, authenticated and scoped.
Agent goals, permissions, action classes and stopping conditions are explicitly bounded.
High-impact, irreversible or financially material actions require approval.
Development, test and production environments are segregated for agent actions.
Unsafe model, prompt, tool or agent changes can be rolled back or isolated quickly.
Fast Company interview with Replit CEO · confidence 92% · last verified 23 Aug 2026
Open underlying source