Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Researchers used MisKnow-Agent to generate convincing but misleading knowledge and tested multiple agents on 5,933 quality-controlled Deep Research examples.

Even limited exposure to credible false information could push final reports toward false conclusions, affecting both open-source and closed-source agents.

Search-based verifier models often detected the misleading evidence in isolation, but that did not stop it from shaping the full research workflow.

Pre- and post-research defenses helped reduce the risk, but they did not eliminate it, highlighting a reliability gap in long-horizon AI research systems.