AI security research / articles
Analysis across cases.
117 Articles · Page 1 of 24
5 matching records on this page
Latest analysis
Northflank, LangGraph and Presenton show different ways credentials may leave their intended storage or diagnostic boundary. Wiz documents what an attacker could then do with two separate Bedrock keys; those are separate observations, not proof of one shared campaign.
AnkiMCP, PDF Reader MCP, Presenton and Volto illustrate different failures: browser origin validation, missing per-request credentials, exposed local files and privileged backend proxies. They require different fixes.
Four controlled studies and one MCP proof of concept show distinct paths from lower-trust text into agent behavior. Their success rates measure different conditions and cannot be added into a universal prompt-injection statistic.
Resource scoping, input sanitation and instruction-channel training reduce different errors. Their security metrics cannot replace authorized-task completion.
Where malicious text is placed, how it is serialized and whether it becomes remembered user intent can matter as much as its wording.
No matches on this page.