Agentic AI
Edera Launches Hypervisor-Level Security to Shield AI Agents From Credential Theft
A new capability-based security model enforces granular access controls at the infrastructure layer, keeping autonomous AI workloads isolated from sensitive API keys and credentials.
InfoQ
Edera has unveiled Object Capability Bounded Authority, a security framework designed to protect autonomous AI agents and Model Context Protocol servers by preventing access to sensitive credentials. The approach, led by Edera co-founder Ariadne Conill, shifts access control enforcement from the application level down to the hypervisor itself—addressing vulnerabilities that arise when AI workloads run in production environments exposed to prompt injection and supply chain attacks.
The Core Problem#
As organisations deploy autonomous coding agents and integrate external tools into production systems, they typically inject API tokens, cloud keys, and administrative credentials directly into runtime environments. If a language model-driven agent is compromised, attackers can extract those secrets and gain lateral movement across internal infrastructure. Traditional container security—relying on boundary isolation and static environment variables—offers limited protection once an attacker penetrates the workload boundary.
How Capability-Based Isolation Works#
The security model decouples the operational authority needed to complete a task from the actual credentials required. Rather than handing an AI agent a personal access token to modify a GitHub repository, the system authorises the agent to perform that specific action without ever exposing the underlying authentication material.
Several runtime components work together to enforce this separation:
- capcurl — Routes API requests through an isolated broker, keeping credentials hidden from the agent
- capsudo — Manages scoped command execution, allowing workloads to run specific privileged commands without standing root access
- SPIFFE and SPIRE integrations — Establish cryptographic workload identity and mutual TLS connections, eliminating reliance on long-lived static secrets
Containing Model Context Protocol Servers#
Model Context Protocol servers frequently serve as high-value attack targets by bridging language models to enterprise databases and tools. By defining explicit object capabilities at the infrastructure layer, teams can restrict an MCP server to specific resources and commands, preventing unchecked access even if the host model is compromised or suffers a prompt injection attack. This approach shrinks the blast radius from environment-wide access down to discrete, task-specific actions.

