The AI Attack Surface: How Agents Raise the Cyber Stakes

The AI Attack Surface article explains that agentic AI tools can be hijacked to subvert an agent’s goals and alter interactions across networks, increasing the risk posed by autonomous systems. According to Accorian AI engineer Nagarjun Rallapalli, these agents are more capable than traditional LLM chatbots, as they can plan, access tools, and set goals, which raises the potential for misuse.

He notes that at Black Hat Middle East & Africa, he will demonstrate how modern agentic AI is vulnerable to a range of attack types, including prompt-based hijacking, time-based attacks, and manipulating inter-agent interactions to compromise entire networks.

The piece highlights a first public example, CVE-2025-53773, involving VS Code and the GitHub Copilot Agent, where a single line could cause the agent to automatically approve tools and potentially fully compromise a developer’s machine, with a demo showing an attacker injecting a prompt to grant permissions. Protecting against these threats emphasises strict access controls, guardrails, and whitelists to limit which APIs agents can call.