AI security research / articles
Analysis across cases.
19 Articles · Page 4 of 4
4 matching records on this page
From the archive
OpenAI, Anthropic, AISI and Google/Irregular exposed different paths from an evaluation objective to real systems. Together they make one old assumption difficult to keep: the sandbox is only one part of the boundary.
The recovery risk is not merely whether synthetic video looks convincing. It is what a successful check authorizes next: in the KZ-CERT account, video verification preceded replacement of the trusted phone number and subsequent access.
A wiki edit, a malicious pull request and a package upload are different actions, but all turn the public internet into persistent state outside the intended task.
Five 2026 records show how workflow nodes, agent skills, MCP servers and memory plugins can become private-data attack paths once they inherit legitimate authority. The evidence does not support blaming free or open-source software as a class: the documented failures include malicious third-party packages, compromised legitimate releases, mutable remote semantics and a separate product authorization bug.
No matches on this page.