https://news.ycombinator.com/item?id=47196883
https://news.ycombinator.com/item?id=47134473
https://news.ycombinator.com/item?id=47147764
https://news.ycombinator.com/item?id=47141321
Besides that.. Agents reporting task completion while the system state says otherwise is predictable once you think about it. Next-token prediction optimizes for plausible outputs, not ground truth.
Quoting from the abstract:
"We report an exploratory red-teaming study of autonomous language-model–powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions."
"Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover."
...a completely unsurprising result, but it's nice to see published experiments.
Any agent system using current LLMs is likely to exhibit undesirable traits that derive from the training data.
The research areas of model alignment and safety are attempting to address this fundamental problem - and have yet to solve it convincingly.
Problems like emergent misalignment can make things even worse.
