Darktrace Launches SECURE AI to Counter Unmanaged Multi-Model Proliferation and Agent Risks

Darktrace announced the general availability of SECURE AI, bringing its core behavioral detection architecture directly to enterprise AI usage and autonomous agent ecosystems. The platform provides real-time security inspection across prominent foundation model providers and cloud platforms, including Amazon Web Services, Microsoft, OpenAI, and Anthropic. SECURE AI evaluates active sessions across interfaces like ChatGPT Enterprise, Claude, Microsoft Copilot, and Amazon Bedrock, scanning for indirect prompt injection, jailbreak attempts, and accidental sensitive data exposure, and providing analysts with risk-ranked session observability. This release tackles a critical pain point that platform engineers, SecOps teams, and enterprise AI architects face: the explosive proliferation of unmanaged AI endpoints and rogue agents. According to Darktrace telemetry from roughly 8,200 deployments, more than 80% of monitored accounts used generative AI services, with organizations engaging an average of five distinct AI providers simultaneously. Early findings highlighted severe governance blind spots, including low-code environments running dozens of untracked agents and contractor teams deploying unsanctioned models without security oversight. Real-time behavioral anomaly detection establishes a needed control plane for enterprises operating complex, multi-model architectures. This development fits into the broader enterprise transition from simple conversational assistants to distributed, multi-agent systems. Over the past year, enterprise AI design has shifted toward autonomous workflows capable of invoking APIs, modifying data pipelines, and adjusting production configurations. However, traditional perimeter firewalls and static access management rules fail to account for non-deterministic AI actions, indirect prompt injections, or internal reconnaissance by compromised agents. Integrating continuous behavioral monitoring alongside cloud-native AI gateways represents the standardizing of runtime security for generative infrastructure. In practice, engineering and security leads should treat AI agent monitoring as a core tier of their observability and DevSecOps pipelines. Relying solely on static corporate acceptable-use policies or pre-deployment evaluations is insufficient when autonomous agents handle real-time customer data and backend execution rights. Organizations must implement programmatic guardrails and runtime telemetry to surface anomalous query volumes, data exfiltration attempts, and unauthorized agent spin-ups before integrating autonomous agents into critical production operations.