Edera, a company that specializes in securing software containers, has announced a new approach to securing autonomous AI agents. Called Object Capability Bounded Authority, it enables enterprises to grant agents permission to perform specific actions without exposing the credentials behind them.
Companies which use AI coding agents, Model Context Protocol (MCP) servers, and other autonomous workloads face a fundamental security dilemma. Such agents need access to company files, repositories, or APIs to perform useful work, but handing over login credentials such as GitHub access tokens creates significant risk if they are compromised or manipulated.
Edera’s Object Capability (OCAP) Bounded Authority addresses this problem by separating what an agent is allowed to do from the credentials required to do it. Rather than placing sensitive tokens inside an agent’s execution environment, Edera enables platform and security teams to grant narrowly defined capabilities that are enforced outside the workload. This practical implementation of object capabilities is enforced by the underlying hardware, the hypervisor. For example, an AI coding agent can be authorized to work within a single GitHub repository without possessing a GitHub token.
OCAP Bounded Authority also limits the scope of GitHub to a specific set of actions, even if the token has broader access. With these limitations, if an AI agent is compromised or goes rogue, the damage is strictly contained. Because the agent never held any credentials in the first place, there are no secrets for an attacker to steal, and the agent’s actions remain locked down to exactly what it was permitted to do.
Commenting on this new approach, Ariadne Conill, co-founder of Edera said:
“The problem with giving AI agents autonomy is that we’ve been giving them authority they don’t actually need. An agent might need permission to update one repository or execute one command, but that shouldn’t require handing it credentials that open the door to an entire environment. Object capabilities let us separate the action from the authority behind it, so agents can do useful work without becoming a security liability.”
The implications extend beyond coding agents to MCP servers, which connect AI models to enterprise tools, commands and APIs. Because these servers frequently operate with access to sensitive credentials and infrastructure, they can become high-value targets for attackers. With object capabilities, organizations can define precisely which commands an MCP server may execute and which external resources it may access, without granting unrestricted access to the underlying systems.
The approach also complements Edera’s existing hardware-enforced workload isolation. While isolation limits the damage an exploited workload can cause to neighboring infrastructure, object capabilities restrict the authority the workload can exercise in the first place.
Conill added:
“The future of agent security can’t depend on predicting every bad decision an AI system might make. We need infrastructure that assumes agents can be compromised and ensures that even when something goes wrong, their authority remains limited to exactly what they were permitted to do.”
Edera is developing Bounded Authority as part of its broader approach to securing AI agents on Kubernetes. Initial functionality of Edera’s Agent Runtime includes scoped command execution and workload identity, with credential-brokered API access and deeper integrations currently under development.
Click Here For The Original Source.
——————————————————–
