AG05:2026 — Unexpected Code Execution

Slide 9 · Overview

Six vulnerability types — six different mechanisms by which agent-generated code does more than it should

Type 1

Sandbox escape

Code that breaks out of its intended execution container to gain access to the underlying host or adjacent systems

Type 2

Unintended filesystem operations

Code that reads, modifies, deletes, or exfiltrates files outside the scope of the task — including production data and system configuration

Type 3

Network-based exfiltration

Code that opens outbound network connections to send data — credentials, files, query results — to external endpoints

Type 4

Privilege escalation via code

Code that exploits the agent's execution environment or available tools to gain access or permissions beyond what was granted

Type 5

Backdoor insertion

Generated application code that contains hidden malicious functionality — hardcoded credentials, reverse shells, hidden endpoints — committed and deployed before anyone reviews it

Type 6

Resource exhaustion

Recursive, unbounded, or infinitely-looping code that consumes CPU, memory, disk, or API quota until services fail

Common thread

All six types share the same underlying gap: the space between what the instruction meant and what the code did — with no human review between generation and execution to catch the difference.

← Back Type 1 — Sandbox Escape →