ls writeups/ai/

AI & LLM security

Attacking AI agents rather than hosts: tool-level RCE, MCP abuse, trust boundaries between agents, and the business-logic failures that appear once an LLM can call tools. The first set comes from the Arkx AI Red-Team CTF (Advanced and Expert tracks).

Advanced · Level 5
Bad Robot

A calculator agent that runs your input through eval() — bypassing an input denylist and a numeric-only output filter by encoding a file as one big integer.

CriticalPython eval() code injection via an agent tool
Advanced · Level 7
Master of Resources

The UI is 'down', but the backing MCP server isn't — its only guard is an Origin check that a non-browser client bypasses by sending no Origin at all.

HighMCP resources/tools exposed behind an Origin-only (DNS-rebinding) guard
Advanced · Level 9
Blind Network Ninja

A connectivity tool drops host straight into a shell and only says Online/Offline — so read the flag out-of-band through the server's own access logs.

CriticalBlind OS command injection with out-of-band exfiltration
Expert · Level 6
The Shape Shifter

A config processor parses your YAML with yaml.load() and prints only the keys — so put the result of code execution in the key position.

CriticalUnsafe YAML deserialization (yaml.load) → RCE
Expert · Level 7
The Octopus

Three agents, one trust boundary. Agent C authorizes on clearance state that Agent B writes from user-controlled file content — poison the cache, become admin.

HighCross-agent trust-boundary failure → privilege escalation via cache poisoning
Expert · Level 8
The Heist

Balance is checked at order creation but only deducted at execution — create both orders first, then execute both, and spend the same $1000 twice.

HighTOCTOU race condition / double-spend (CWE-367)