{
 "accessed": "2026-10-12",
 "schema": "CAIN42/EVOLUTION15/RESEARCH-MATRIX/v1",
 "sources": [
  {
   "authors": "Jiang, Wang, Liang, Wang",
   "contribution": "644 cases, 28 environments, 5 long-horizon attack types (intent hijacking, tool chaining, task injection, objective drift, memory poisoning)",
   "implication": "taxonomy adopted; their LLM agents not run here (no model in loop)",
   "overlap": "our attack families cover the same 5 categories in a deterministic simulator",
   "title": "AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks",
   "url": "https://proceedings.mlr.press/v306/jiang26as.html",
   "venue": "ICML 2026, PMLR 306:53162-53180 (peer reviewed); arXiv 2602.16901"
  },
  {
   "authors": "Kaptein, Khan, Podstavnychy",
   "contribution": "policies as functions of agent identity, partial path, next action and organizational state",
   "implication": "path-level governance is published prior art; our policies are deterministic, not probability estimates",
   "overlap": "the runtime evaluates identity + delegation state + plan history + memory provenance at each step",
   "title": "Runtime Governance for AI Agents: Policies on Paths",
   "url": "https://arxiv.org/abs/2603.16586",
   "venue": "arXiv 2603.16586 (preprint, 2026-03-17)"
  },
  {
   "contribution": "trace policies (temporal, argument, value-flow) compiled to SMT over finite traces",
   "implication": "prior art for trace/value-flow enforcement",
   "overlap": "taint from untrusted memory is a value-flow condition",
   "title": "VIGIL: Runtime Enforcement of Behavioral Specifications in AI Agent Skills",
   "url": "https://arxiv.org/abs/2606.26524",
   "venue": "arXiv 2606.26524 (preprint, 2026-06-25)"
  },
  {
   "contribution": "only the abstract could be checked: runtime controllers against prompt injection propagating through workflows",
   "implication": "treated as prior art; details not relied on",
   "overlap": "unverified",
   "title": "SafeAgent: A Runtime Protection Architecture for Agentic Systems",
   "url": "https://arxiv.org/abs/2604.17562",
   "venue": "arXiv 2604.17562 (preprint)"
  },
  {
   "contribution": "NOT READ in this session",
   "implication": "memory evaluation here is security-only; task-performance memory benchmarks not run",
   "title": "AMA-Bench; Benchmarking Agent Memory in Interdependent Multi-Session Tasks",
   "url": "https://proceedings.mlr.press/v306/zhao26bs.html",
   "venue": "PMLR 306 (listed by the user)"
  },
  {
   "contribution": "NOT READ in this session",
   "implication": "listed as unreviewed prior art",
   "title": "Governance by Construction for Generalist Agents (IBM Research)",
   "url": "https://research.ibm.com/publications/governance-by-construction-for-generalist-agents--1"
  },
  {
   "contribution": "industry standards, open protocols, agent security and identity research; NCCoE identity & authorization concept paper (not final)",
   "implication": "no conformance claimed; the runtime's identity/authorization separation is aligned in spirit only",
   "title": "NIST AI Agent Standards Initiative (CAISI)",
   "url": "https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative",
   "venue": "NIST, launched 2026-02"
  }
 ]
}
