Name-Resolution Confusability / Homoglyph Tool & Recipient Spoofing Check

Your agent resolved a confusable imposter of the target it meant to call.

A recognized real-world attack class no other surface grades: when an agent resolves a referenced tool / recipient / domain / endpoint / package NAME and a visually-confusable lookalike — a Cyrillic 'р' in striрe.com, a zero-width-padded transfеr_funds, an IDN/punycode twin, a case-fold / whitespace collision — it silently routes a PRIVILEGED action to the WRONG, attacker-controlled target. It is a DISTINCT trust boundary from every shipped surface: obfuscation / metamorphic grade disguised INPUT the agent reads; tool-pivot grades being STEERED to a REAL privileged tool; tool-meta grades a tool's own metadata; supply-chain grades package integrity not name confusability at call time; cache grades response bleed. NR.1 is the agent CALLING a confusable IMPOSTER of the intended target. AgentSmack reduces each intended and resolved name to a Unicode skeleton (NFKC + homoglyph-fold against the shared Cyrillic table + invisible-char-strip + case-fold + whitespace-strip), compares them, and fires a hard blocker when a resolved name diverges from intent or a privileged action lands on a confusable target. Ground-truth-first: a byte-identical, trusted- source resolution scores a clean 100; a missing intended name is worst-case. Raw full names are never stored — only a sha-prefix + skeleton labels + closed enums + a redacted excerpt. Fill with a sample to watch the intended → resolved → action lineage light up with no live infra.