[AWS Summit Seoul 2026] Irresponsible AI Agents, Who's in Charge

As AI technology expands into the physical world, the risks posed by the autonomy of AI agents are growing. By 2028, 15% of corporate decisions are expected to be made autonomously by AI, but many companies are currently exposed to security gaps.

The major problems that arise during AI adoption are as follows.

• Lack of access control (97%): Uploading source code and other materials to LLMs indiscriminately without proper controls
• Lack of guardrails (63%): Insufficient rules or regulations when accessing multiple systems
• Confidential information leaks (60%): Data leakage risk due to security vulnerabilities
• Additional cost spending: Massive losses from uncontrolled shadow AI and security incidents

Prompt Engineering, which simply improves the quality of responses, alone cannot prevent unexpected behaviors such as hallucination or prompt injection. Therefore, Harness Engineering, which designs AI to operate only within defined safe zones, is essential.

The key implementation methods of Harness Engineering are as follows.

• Minimizing permissions (Action/Skill Control): Restricting AI to only call predefined APIs so it cannot arbitrarily modify systems
• Inserting a human approval step (Human-in-the-loop): Enforcing a pipeline that requires operator approval right before final execution
• Multi-agent cross-validation (Chaining & Evaluation): Designing a structure where an independent evaluation AI verifies the results of the executing AI