From the source
Baseten says it has layered new security controls across its inference platform in response to attackers who now use agent swarms to find exploits and exfiltrate data within minutes.
- The company says AI assistance has let threat actors move from AI-assisted attacks to fully autonomous AI-led attacks, citing Google's Q4 AI Threat Tracker, and that LLM inference has become a prime target because attacks aim at model weights and GPU capacity.
- Baseten describes defenses it has added: hardened containers, mutually authenticated request paths, isolated customer workloads, agent-based detection that flags malicious activity in seconds, and agentic red teaming against its own code alongside third-party penetration testing.
- Customers keep control of identity through their own IdP with fine-grained roles, use scoped short-lived tokens instead of stored keys, and can encrypt served weights with keys in their own cloud KMS and revoke Baseten's access at any time.
- Network boundaries can be set with ingress and egress allowlists, single-tenant deployments can use private connectivity, inference runs only in approved regions with zero data retention, and audit logs are exportable to customers' own detection tools.
- Baseten says it will publish more about its security programs in the coming weeks, covering API key hygiene, container and network security, identity and access management, and observability, detection and response.






