section // posts
/topics/ai-agent-security

AI and agent security

Production AI systems assessed through their data, deployment identities, tools, and real-world authority—not prompt lists alone.

1 posts2 CVE analysesAll topics
Full topic archive

Every matching record.

Methods and named-vulnerability research remain visually and editorially separate.

Field notes 1

AI Security · Nov 9, 2026

Breaking AI Agent Boundaries: From Prompt Injection to Remote Code Execution via Model Context Protocol (MCP)

How untrusted context in modern AI pipelines forces autonomous agents to execute arbitrary operating system commands through Model Context Protocol (MCP) server tools: indirect injection, ambient authority, and deterministic execution brokers.

14 min read ↗

CVE research 2

CVE-2026-125330 · Lab Reproduced

CVE-2026-125330: Cross-Workspace Session Access with a HarnessRouter Workspace Key

Root-cause analysis of CVE-2026-125330: HarnessRouter narrowed the session list to an API key's workspace, but the routes that take a session or response id checked only the organization. A workspace key that knew another workspace's session id could read, overwrite, publish and delete that session. Fixed in 0.31.7.

17 min read ↗CVE-2026-107716 · Lab Reproduced

CVE-2026-107716: Symlink Traversal in banks Prompt Registry

Root-cause analysis of CVE-2026-107716: DirectoryPromptRegistry in banks followed symbolic links out of its prompt directory, so a link could make it read an arbitrary file through _scan() and get(), or overwrite one through set() and index.json.

10 min read ↗