Safety and Trust

In the last module, the agent learned to remember and retrieve. An agent that remembers can act on what it remembers, and it can be taught things that are not true. In this module, you learn why a prompt is not a security boundary and how tools can poison memory, then optionally watch guards and provenance stop a real attack.

By the end of this module, you will be able to:

• Explain why a prompt is not a security boundary
• Describe how a tool result can poison memory
• Trace a fact to its source with source, trust and a link to the item it came from
• Optionally, block tool calls with an approval gate and a recipient allowlist, and run the poisoning attack with provenance on