Inside the lethal trifecta: Blast radius reduction in AI agent deployments

This article warns that AI agents which read untrusted content, access sensitive systems, and communicate externally are vulnerable to indirect prompt injection; it surveys evidence (research demos and a Google Common Crawl study) and prescribes seven pragmatic patterns—agent sandboxing, credential isolation, sealed tool endpoints, egress restriction and monitoring, EDR coverage, human-gated approvals, and injection-propagation boundaries—to assume-breach and contain blast radius while tooling matures.