HybridClaw’s new Learning mode records which approvals an action would need while letting the action actually happen, including protected-approval cases. The developer says authentication, sandbox limits, active safety blocks and explicit denials still apply. The release leaves the default approval mode unchanged and adds an administrator view for examining flagged actions and their outcomes.
HybridAI has introduced a new way to examine approval requirements in its HybridClaw agent software. The optional Learning setting lets actions proceed while recording the permission requests they would normally produce. This includes operations ordinarily requiring protected approval. The release notes explicitly warn that the setting changes real systems: it does not merely simulate what an agent might do.[1]
The developer says identity checks, restrictions imposed by the sandbox and enabled safety blocks continue to constrain actions. An explicit denial still applies. Auto remains the default approval setting, alongside active safety blocks. Administrators can inspect flagged operations and outcomes in an audit view. The security interface also brings together approval defaults for future conversations and changes to individual conversations; configuration changes are audited and outdated edits are refused.[1]
HybridClaw 0.42.0 also tightens replacement of existing files: custom callers must read the whole file and supply expected contents identical to its current state. If that state changes, another read is required. When a conversation is interrupted, completed tool results and messages are kept, while effects whose status remains uncertain are flagged. These are changes described by the developer, without an independent security assessment establishing their effectiveness.[1]
