1.1.1

🐕 AgentHound Changelog

1.1.1 — Evidence Precision and Reliability

Added

• Instruction findings now preserve bounded matched excerpts, source positions, scope, and file metadata from the scan artifact through the dashboard and copied report.
• A medium-severity INSTRUCTION_SIGNAL finding separates reviewable standalone or deep-scope signals from high-confidence poisoning in active instruction scope.

Fixed

• Instruction-file classification now requires deterministic override, identity, hidden-content, decoded-payload, or sensitive-outbound semantics instead of promoting ordinary instruction language and documentation examples.
• Concrete credentials discovered in multiple agent configurations now share one value_hash-based identity with deterministic provenance and authentication hints.
• Positional scan targets now reject malformed hostname syntax before any resolver or network work begins.
• The installer now validates before promotion and atomically restores the previous binary when installed-path startup validation fails.
• Legitimate documentation URLs no longer look like Base64 prompt-injection payloads or, by themselves, outbound-network capabilities.
• Active planning avoids redundant credential presentation when the exact MCP resource or A2A card/probe surface has already succeeded anonymously.
• Repeated service collection now keeps protected Open WebUI posture stable across credential attempts and gives each distinct LiteLLM master key a contextual value_hash identity.