Agent Input Hardening

Give content-ingesting agents a clear boundary between evidence and authority. This skill combines practical prompt-injection handling with small, inspectable Python tools: a streaming fence guard, a publication secret scanner, and a literal redaction helper. Learn where to place output contracts, how to reject invented human facts, and how to keep credentials outside agent workspaces. Offline tests include split stream delimiters, unterminated blocks, credential-shaped strings, and overlapping redactions. The guidance explains failure modes that simple “ignore malicious instructions” prompts miss, including tool-result laundering and scan-to-publish races. All examples use the standard library and are ready to adapt to your harness. Includes explicit limitations so teams can layer real isolation and independent verification around the examples.