New benchmark tests AI agent safety against multi-step prompt injection attacks

Researchers have introduced StepJack, a new benchmark designed to test the safety of computer-use agents (CUAs) against multi-step indirect prompt injection attacks. These attacks involve distributing adversarial instructions across a chain of web pages, making them appear innocuous individually. Evaluations on the StepJack benchmark, which includes 480 test examples, revealed that multi-step attacks significantly increased the success rate for several state-of-the-art CUAs, including GPT-5.4 Mini, by up to 31.2 percentage points. AI

IMPACT This research highlights a new vulnerability in AI agents, potentially requiring developers to implement more robust safety measures against sophisticated prompt injection techniques.

RANK_REASON The cluster describes a new academic paper introducing a novel benchmark and attack class for AI safety research. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Computer Use Agents
• EvoCUA-32B
• GPT 5.4 Mini
• indirect prompt injection
• multi-step indirect prompt injection
• StepJack

• safety
• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →