Policy dependency / Stack layer
Hierarchical Ransomware Agents Escalate to Dynamic and Memory Analysis Only on Specialist Disagreement
arXiv 2609.04820
Policy dependency / Stack layer
83.3% of 281 open-source AI contribution policies permit AI code, but 67.3% demand a high level of human involvement
arXiv
Policy dependency / Threat pattern
None of five agent-memory systems actually enforce revocation at retrieval time
arXiv
Stack layer / Threat pattern
A Review Through 31 August 2026 Finds Agent Action Surfaces Expanded Far More Than Reliable Completion Did
arXiv 2609.04894
Stack layer / Threat pattern
16% of 3,171 public agent-harness setups carry a confirmed security defect, and 3.8% ship a skill that pre-approves your shell
arXiv
Stack layer / Threat pattern
A capability-scoped harness cut prompt-injection execution from 33-47/75 runs to 3/75 without asking the model to spot malicious text
arXiv 2609.08371
Stack layer / Threat pattern
The authority a coding agent needs often lives outside its workspace, and augmenting the planner's view does not fix it
arXiv
Stack layer / Threat pattern
Handing an agent a 128 MB and 10-second execution contract made generated code up to 3.1x faster and cut peak memory in 13 of 14 comparisons
arXiv