Instinct blamed the delivery side for a stranger's document appearing in one user's thread

An assistant asked whose name is on a document reads its context window. Asked how the document got into that window, it reads the same window and writes something that fits. Instinct said it had checked its records and found that Pritak's message contained only text ^[5]. It said a Gerber image had crossed into his conversation "somewhere on the delivery side" and accepted responsibility for using it ^[6].

That is the only account of the routing layer in the screenshots, which runtimewire.com describes from posts Pritak made on X on September 21st ^[1]. Instinct's own terms of service warn that outputs may contain incorrect or incomplete information and that records of actions available through Instinct may be inaccurate ^[9]. Instinct offered to report the incident and then wrote the report itself ^[7]. The draft said a photo the user never sent had appeared in one of his text conversations and concluded that a private document had reached the wrong conversation ^[8]. The same model produced that report and the wrong middle name ^[4].

What separates the two readings sits on the server: attachment IDs scoped to one conversation, object ACLs keyed to a user, a log line naming the object the model actually read. None of that appears in a screenshot. runtimewire.com's account does not include a comment from Shinn ^[23]. Replies speculated about "mishandling of per-user file workspaces", and runtimewire.com notes those are user interpretations, with no confirmed diagnosis of Instinct's architecture ^[15].

The permissions are what make the ambiguity expensive. Instinct's privacy policy says the assistant may process documents, private messages, emails, account credentials, payment information and health-related information when users grant the corresponding permissions ^[16]. Its terms authorize the service to access, copy, collect and index information from connected accounts ^[17]. Shinn built Instinct to act through the applications people already use, and users connect outside accounts and authorize it to complete tasks including reservations and purchases ^[18]. Instinct told Pritak nothing had been filed and that it would abandon the $574 claim ^[7].

The reliability tally in the replies has a smaller version of the same problem. @peepeedog said that across 29 questions Instinct was "wrong or misleading" 10 times and apologized or walked back an answer 11 times ^[11], which is about 34 percent on the first count ^[19]. @SSShken asked whether competing assistants received the same questions in the same order and whether the walk-backs overlapped with the incorrect answers ^[12]. For the tally to say anything comparative, the same 29 prompts would have to run against the other assistant in the same order, with walk-backs counted in their own column. The overlap also sets the size of the claim: if the two sets are disjoint they touch 21 of 29 responses, about 72 percent ^[20]; if they overlap completely, 11.

Clayton Bradshaw replied that a similar episode had "happened to me too, then Instinct denied it" ^[13]. One reply put it plainly, writing that "Data leak or hallucination?" was "not a reassuring multiple-choice question" ^[14].