Task-graph contextual integrity limits multi-agent privacy leaks
On MAPLE-Bench's 144 cases, TG-CI cut mean unauthorized disclosures 85.8-94.1% while keeping 93.6-96.1% of baseline task accuracy.
3 stories, newest first · one card per story
On MAPLE-Bench's 144 cases, TG-CI cut mean unauthorized disclosures 85.8-94.1% while keeping 93.6-96.1% of baseline task accuracy.
On a held-out set, the best existing tools left a whole value visible in 61% of Russian and 58% of UK English documents.
Prompt embeddings re-identify 50-75% of SWE-Chat users behind anonymizing LLM proxies.
Today 1.1K readers · 1.3K views · 178 reads · all time 203K views