01 · Agent attacks
A tricked agent never gets to act
Coding agents read READMEs, emails and web pages. One hidden instruction can turn an agent against you. AIDR checks every command, file write, web request and MCP call (an agent’s link to other systems) before it runs. The harmful step stops; normal work carries on.
- Hidden instructions in READMEs, documents and emails
- Poisoned tool output hidden before the AI reads it
- MCP servers and tools that turn malicious after approval
- Multi-step attacks judged as one chain
- Key theft, fake packages and disguised destructive commands