Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning
- Published
- Record updated
Summary
This paper systematizes security research on deep research agents across four stages: retrieval, planning, evidence integration, and report generation. Coding 34 primary studies, it finds only 26.5% address the planning stage and none run an end-to-end poisoning evaluation on commercial products. The authors introduce the agenda seeding attack, which adds a fictitious evaluation metric without any instructions, and it makes the attacker's target the top pick in 16% to 28% of 3,520 sandboxed runs, while GEO and prompt injection are nearly ineffective.
Mitigation
The source states that provenance tagging preserves utility but can be bypassed by multi-site bait, and that injection detection and query anchoring fail against agenda seeding while verification prompting hurts utility. It proposes eight research directions but does not present a fix for agenda seeding.
Related items
- MediumCVE-2026-130358: AdminNeo before 5.9.1 contains a missing authentication vulnerability in the ClaudeSqlPlugin, GeminiSqlPlugin, and…Similar attack · NVD/CVE Database
- HighGHSA-8vf2-hxg5-crqh: Kubeshark: MCP `start_kubeshark` argument injection via `pod_regex` → image redirect → cluster takeoverSimilar attack · GitHub Advisory Database
- MediumCVE-2026-126681: AWS Labs Timestream for InfluxDB MCP Server exposes API tokenSimilar attack · NVD/CVE Database
- HighCVE-2026-126681 - Improper restriction of intended endpoints in AWS Labs Timestream for InfluxDB MCP ServerSimilar attack · AWS Security Bulletins
- InfoRisk-certified prompt optimization for LLM-as-a-judge guardrailsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)