Pi Agent Platform v1.44.3

Prompt injection cannot quietly read the system. Text in a page, file, memory or helper answer can no longer send data to an unknown site, read credentials in Bypass or steer the agent without the member seeing it.

Security

• web_fetch reads documentation and registry sites freely; another site is asked about once per conversation, and an address there with a query or fragment every time.
• Helper answers and team memory results reach the agent inside a data fence it cannot close; the agent never shows its configuration or credentials or asks for a secret on their word.
• In Bypass, run_on_host still asks when a command reaches credentials: ~/.ssh, ~/.pi, .env, auth.json, keys, the Keychain or the environment.