Deep research System Card

Source: OpenAI Date: 2025-02-25 URL: https://openai.com/index/deep-research-system-card

Summary

Summary

The Deep Research system card documents safety evaluations and risk mitigations for OpenAI’s autonomous research agent feature. Key risks evaluated include: generating comprehensive harmful information through systematic research (the autonomy amplifies output quality for dual-use queries), potential for bias and misinformation in synthesized reports, and the appropriate scope of autonomous web browsing without user confirmation.

Implications

Safety/agentic thread. Deep Research’s system card is important because it documents the first class of agent safety challenges that differ structurally from chatbot safety: an autonomous research agent can assemble harmful information more comprehensively than a single-shot query by conducting systematic research across multiple sources. The card reflects OpenAI’s attempt to apply the safety evaluation framework to agentic products, but the evaluation methodology for agents is less developed than for language model outputs. The February 2025 timing is early in the agentic product era, and the system card sets expectations that were tested as the product matured and its capabilities expanded.