AI security

Prompt injection, agents and MCP servers: the new attack surface of web apps.

Computer screen seen at an angle, showing colorful code in a dark room

AI security9 Sept 2026 · 6 min read

Prompt injection isn't a bug you patch once

Researcher RyotaK ended up cataloguing around fifty separate ways to bypass Claude Code's permission system. The most critical one let an attacker steal CI/CD secrets in January 2026, starting from a single GitHub issue.

Backlit keyboard in the dark, close-up on the command and option keys

AI security19 Aug 2026 · 6 min read

The AI assistant blind spot: who's actually allowed to touch your infra

CVE-2026-22708: in Cursor, a plain README or an issue comment was enough to mutate an environment variable and hijack a command the allowlist had already approved, something as ordinary as git branch.