Aim Intelligence Outlines Context-Aware Agent Security

Aim Intelligence Outlines Context-Aware Agent Security
Monday, November 16, 2026
  • •Aim Intelligence says AI agent actions should be governed in the context of the full workflow.
  • •Its security framework identifies Shadow AI and connected data and tools, while tracking permissions and execution.
  • •Its Starport platform combines real-time guardrails, execution tracking, document export checks, and communications blocking.

Aim Intelligence says companies should inspect how enterprise AI agents use data and tools, monitor interactions between agents, and control their actions. Hanul Lee, head of the company’s Defense Security division, said companies should assess whether actions are appropriate in the context of the full workflow, not only whether each action falls within its permitted scope. He said control technology is developing more slowly than model performance. Risks include people giving harmful instructions deliberately or by mistake, and agents exceeding their permitted scope or bypassing safeguards while pursuing their goals.

Aim Intelligence proposed a security framework covering AI asset discovery, execution tracking, real-time control of permissions and actions, and incident response. Companies should first find “Shadow AI” (AI tools used outside official company oversight), such as external services employees access through personal accounts and internally developed tools. They should identify not only models but also connected external tools, data, memory, and Model Context Protocol (MCP) servers. One customer found more agents than its executives had expected after applying the framework. The company grouped AI entry points into three categories: employees’ use of external services, internal platform development, and customer-facing services. Each carries different risks, including confidential data leaks, data permissions and misuse, malicious prompts, and control of unauthorized actions. Companies should update their inventories when agents or tool connections are added.

When agents collaborate, a higher-level agent should not pass unnecessary permissions to lower-level agents. Lee said each agent should have only the minimum permissions needed for its assigned work, and higher-level agents’ authentication tokens or passwords should not be passed directly to lower-level agents. Even when an agent has permission, whether it should call a tool for a particular task requires a separate judgment. Aim Intelligence emphasized observability (recording and analyzing execution) to log and analyze inputs, retrieved data, tool-call timing, and links between tasks. It said companies should examine the context, including the original instruction and the reason for an API call.

The company said it combines static security, which blocks access to prohibited networks or systems under preset rules, with dynamic security, which assesses actions based on the current task and earlier steps. Analyzing every record each time can increase processing time and computing costs, so checks are divided into stages according to risk. The first stage detects indirect prompt injection (hidden instructions mistaken for legitimate commands), which can be embedded in external data, tool or MCP server descriptions, skills, and memory. Clearly prohibited access and tool execution are blocked; calls requiring further assessment are checked against prior execution records and workflow context.

Aim Intelligence developed Starport, an AI security platform that applies policies to inputs and outputs, retrieval and conversations, and tool execution. Its Guardian component uses real-time guardrails to check for prompt injection, sensitive information leaks, policy violations, and tool and MCP connections. Opticon records and tracks execution flows, while Stratum classifies documents sent externally and checks them before export. These functions can be applied between employees’ PCs and AI services, in enterprise AI applications, APIs, and proxy servers. Starport also supports a kill switch to block AI communications by project or organization.

The company operates Stinger, a red-teaming tool that tests AI models and systems using attack scenarios. It uses attack types found in testing to improve Guardian’s defenses and detection. Aim Intelligence said purple teaming, which tests attacks and defenses together, helps incorporate new attack techniques into defensive measures. Pre-launch vulnerability checks do not eliminate risks during operation: new risks can arise when connected tools, data, or execution scope change, requiring ongoing testing and policy updates.

Lee said requiring people to give final approval for every task is not a sufficient solution. A high volume of approval requests can lead people to approve without reviewing the background and impact, so companies should decide which tasks require approval, when approval is needed, and who should provide it. Companies should first define important tasks, distinguish work delegated to agents from work requiring human approval, and give approvers the information needed to make a decision. Aim Intelligence said financial firms tend to seek regulatory compliance features such as data loss prevention and audit records, while technology companies with active AI use tend to seek more controls for operational risks during execution.

Aim Intelligence said many South Korean companies have not y

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox