Raw Corpus Synthesis 2026-07-02
Corpus
Batch ingest covered 255 markdown raw sources: 185 papers, 36 whitepapers/standards/guidance documents, and 34 news or incident captures.
Main Clusters
- Agent Security and Tool Abuse
- Prompt Injection and Context Attacks
- Memory Poisoning and Agent State
- RAG and Retrieval Security
- MCP and Agent Protocol Security
- Agent Identity and Authorization
- AI Cybersecurity Operations
- Evaluation Benchmarks for AI Security
- Model Extraction and Privacy Leakage
- AI Security Governance and Standards
Synthesis
The raw corpus is strongly weighted toward agentic AI security. Recurring themes include prompt injection in tool-integrated environments, memory poisoning, MCP/tool poisoning, RAG poisoning, runtime monitoring, agent identity, benchmark construction, and governance frameworks. The corpus contains many benchmark and defense papers, but the core unresolved issue is whether benchmark performance and proposed controls transfer to production deployments and real incidents.
Evidence Caveat
This is a batch-level synthesis. Individual claims should be stabilized by promoting representative primary sources into source-specific notes and updating claim pages with precise evidence levels.