A new benchmark shows enterprise LLM agents violate contextual integrity at rates of 15.8-50.9% with leakage up to 26.7%, and higher task performance correlates with more privacy breaches that model scaling does not fix.
The adversarial instruction is used to test whether the Copilot can decide what information is appropriate to share when executing the instruction
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
CI-Work: Benchmarking Contextual Integrity in Enterprise LLM Agents
A new benchmark shows enterprise LLM agents violate contextual integrity at rates of 15.8-50.9% with leakage up to 26.7%, and higher task performance correlates with more privacy breaches that model scaling does not fix.