Presented at DEF CON 34
AI agents escape their task horizon
Abstract
Frontier Agents (based on OpenAI, Google and Anthropic LLMs) show some propensity to escape their task’s horizon, by exchanging state with past or future instances of themselves, unprompted, to carry forward context and partial contributions towards their goal. Later agents “gain” from the content left by the earlier agent, achieving some continuity, and breaking independence of tasks (in evaluation or production)