Security Gaps in AI 'Harnesses' Expose Enterprises to Systemic Risk

Security researchers at Novee Security have identified critical weaknesses in the software frameworks, known as "harnesses," that surround large language models from major vendors including Google, Anthropic, and OpenAI. These vulnerabilities stem from trust misalignments between different harness components, potentially allowing attackers to bypass security guardrails.

AI harnesses are the software scaffolding that provides models with essential tools, memory, and feedback loops, as well as context management. According to Novee Security, while vendors have implemented standard security checks and sandboxes, the system fails during the "handoffs" between these various components. This creates a gap where security controls are bypassed as data moves from one part of the harness to another.

The Shift to Agentic AI

This vulnerability emerges as enterprises transition from simple prompt-and-response interactions to "agentic AI," where models autonomously execute code and utilize external tools. As the complexity of this surrounding software increases, the interactions between components often lack rigorous trust boundaries. This architectural flaw creates new attack vectors for prompt injection and supply chain exploits, as the "glue" code connecting a model to its tools is often too trusting.

Systemic Industry Risk

The risk is systemic because organizations adopting AI agents implicitly inherit the trust assumptions and potentially insecure code of the vendor's harness. If these boundaries fail, a helpful agent can be transformed into a security liability capable of unauthorized data modification or remote code execution within corporate infrastructure. Elad Meged, a security researcher on the Novee Security founding team, noted that users are often unaware of the volume of code and trust they embed into their systems when adopting an agent.

A Growing Security Gap

These technical failures coincide with a significant lag in corporate defense strategies. A study published in early 2026 reveals a stark disparity in AI adoption: while 80% of companies are currently running AI agents, only 47% have implemented the necessary security controls to manage the associated risks.

What to Watch

As vendors move to patch these handoff failures, the industry must determine if standard sandboxing is sufficient or if a new architecture for trust boundaries is required. While Novee Security has highlighted these structural flaws, the full extent of how many enterprises have been compromised by these specific harness vulnerabilities remains unconfirmed.