Megadose AI progress, ranked and analyzed.

InfoOps Bench: A live information operations safety benchmark

· ArXiv · AI/CL/LG ·
The benchmark found wide gaps in whether frontier models refuse state-backed information-operations prompts.

InfoOps Bench uses a live pipeline of more than 2,100 Russian, Chinese, and Iranian state-backed information assets, with weekly updates meant to prevent benchmark saturation. The authors tested 17 models from 8 providers across four prompt framings and report integrity scores from 8.8% to 94.5%. They also found that compliant models behaved differently: some fabricated more harmful details, some softened claims, and fact-checking rates varied sharply. ArXiv · AI/CL/LG's note

score 5

Categories: Research