Security Intelligence Center

Malicious tools hide harmful instructions in descriptions invisible to users but read by LLMs, enabling data exfiltration and unauthorized actions.

Initially safe tools are updated to include malicious functionality after gaining user trust and installs.

A compromised tool leverages permissions granted to other trusted tools in the same agent session.

Attackers manipulate tool inputs through crafted prompts, causing tools to execute unintended operations.

Attackers chain multiple low-privilege tools together to achieve actions that no single tool should allow.

Tools request more permissions than necessary for their stated functionality, expanding the attack surface.

API keys and tokens stored in plaintext config files or environment variables accessible to all tools.

Tools execute with host-level access instead of being isolated in containers or sandboxed environments.

No cryptographic verification that a tool has not been tampered with between installation and execution.

Insufficient logging of tool actions makes it impossible to detect or investigate security incidents.