PROVISO, introduced in a new arXiv paper, is a provenance-informed framework for action selection in LLM agents. The authors start from a concrete risk: agents act on the external world through tools and third-party skills, so one executed action can cause irreversible damage. They argue that current defenses are limited because they judge a proposed action without tracking whether its content originates from the user.
By contrast, PROVISO incorporates provenance—where the content behind an action came from—into the decision to allow or deny it. This lets the agent distinguish between user-directed instructions and content from third-party skills or other untrusted sources. The paper positions this as a way to close a blind spot in existing safety mechanisms, though the abstract does not provide experimental details.
As the only source reviewed here, there is no independent comparison to note; the findings are presented as the authors' own proposal.