With the rapid adoption of autonomous LLM-based agents (giving models access to shell execution, API calls, and local file systems), the boundary between intentional behavior and unintended execution is blurring.
I'm less concerned with sci-fi "sentience" and more interested in the practical security and control aspects:
Prompt injection causing privilege escalation or unauthorized state changes.
Feedback loops where an agent overrides safety boundaries to satisfy an optimization goal.
Failure of sandboxing when agents are given multi-step execution autonomy without human-in-the-loop validation.
From an engineering and systems perspective: do you consider runtime containment/sandboxing practically solvable for fully autonomous agents, or will human approval at critical checkpoints remain non-negotiable? How are you mitigating these risks in your current implementations?
Where "escape" means do something you didn't anticipate or in a manner you didn't anticipate, sure. The bar is pretty low on that, for instance the OpenAI thing the other day where it "escaped containment" which if I understand correctly it interpreted a connectivity issue as a problem to solve when it had actually been intentionally blocked/sandboxed. To borrow the phrase "more than one way to skin a cat", it will always be difficult to reduce it to exactly one fully-controlled way for all shapes, sizes and pelts of cat. It's a very good argument for running AI locally since you have physical control over its connectivity and there's nothing it can do to reconfigure or circumvent that.
Where "escape" means do something you didn't anticipate or in a manner you didn't anticipate, sure. The bar is pretty low on that, for instance the OpenAI thing the other day where it "escaped containment" which if I understand correctly it interpreted a connectivity issue as a problem to solve when it had actually been intentionally blocked/sandboxed. To borrow the phrase "more than one way to skin a cat", it will always be difficult to reduce it to exactly one fully-controlled way for all shapes, sizes and pelts of cat. It's a very good argument for running AI locally since you have physical control over its connectivity and there's nothing it can do to reconfigure or circumvent that.