← All concepts

agentic tool use safety

1 articles · 3 co-occurring · 0 contradictions · 0 briefs

pkilling processes without explicit authorization suggests model's understanding of 'what's safe to do' isn't clearly established in context—over-literal and over-autonomous behaviors indicate instruc

Related concepts
subagent context isolation 1 multi turn state degradation 1 context window saturation 1
Evidence chain (1 articles, showing 1)
@davis7: Opus 5 feels like a strange hybrid of GPT-5.6-sol and Fable, and I think I re... example_of

pkilling processes without explicit authorization suggests model's understanding of 'what's safe to do' isn't clearly established in context—over-literal and over-autonomous behaviors indicate instruc

query this concept
$ db.articles("agentic-tool-use-safety")
$ db.cooccurrence("agentic-tool-use-safety")
$ db.contradictions("agentic-tool-use-safety")