Current Synthesis
AI security research in this portal is organized around how AI systems fail, how they can be abused, how those failures can be measured, and how defenses change the risk profile.
Corpus Update
The raw source corpus now includes SRC-20260702-raw-papers-batch, SRC-20260702-raw-whitepapers-batch, and SRC-20260702-raw-news-batch, covering 255 markdown captures. See Raw Corpus Synthesis 2026-07-02 for the first batch synthesis.
Current Working Thesis
The most useful AI security knowledge base should track four layers together:
- Threat model: who can act, what they can observe, and what they can change.
- Mechanism: why the model, agent, toolchain, or deployment fails.
- Evidence: papers, incidents, benchmarks, evaluations, and reproducible examples.
- Mitigation: defenses, monitoring, governance, and residual risk.
Batch-Ingested Emphasis
- Agentic systems expand the security boundary to tools, memory, identity, and protocols.
- Prompt injection is especially consequential when it can influence actions, retrieval, code, browsers, or persistent memory.
- Benchmark-heavy evidence needs validation against production incidents and realistic deployments.
- Governance and standards sources provide taxonomies and controls, but technical evidence must still be mapped to threat models.
Open Synthesis Gaps
- Which attacks have strong empirical evidence versus mainly demonstration value?
- Which defenses generalize across model families, deployment settings, and agent architectures?
- Which benchmarks predict real deployment risk?
- How should incident reports update research priorities?