BoundSec is a defensive security-research tool. It is designed to test AI agent systems you own or are explicitly authorised to assess.

• The Agent Gym is entirely self-contained; its "secrets" are planted canaries with no real value, and it makes no network calls.
• The live-model adapter wraps a chat model in a tool-calling agent whose tools are sandboxed — a dangerous tool call is recorded, not executed. BoundSec observes the model's choice, it does not run rm -rf.
• Do not point BoundSec at systems you do not have permission to test. Unauthorised testing may violate law (CFAA and equivalents).

Please open a private security advisory or email the maintainer rather than filing a public issue. We aim to acknowledge within a few days.