Instinct: Shinn Discusses an Hallucination, No Data Leak

A user believed they saw someone else's photo appear in their chat with Instinct. Noah Shinn, founder of the invitation-only AI assistant, claims it was not a data breach but rather a hallucination, highlighting technical safeguards. He states that he is working on an active detection layer, while his startup recently raised $250 million and Meta is also pushing its personal agents.

The Agent Market Heats Up, Amid Funding Rounds and New Launches

Instinct, an AI assistant available only by invitation, generated significant interest in Silicon Valley this summer. Noah Shinn's startup, Spear Street, raised $250 million, achieving a valuation of $2.5 billion. These personal agents are integrated into daily tasks, such as managing calls to insurance companies or booking movie tickets. Meta launched its own agent, Muse, on September 8, which quickly rose to the top of the most downloaded apps on Apple and Android devices. More broadly, personal AI assistants are experiencing growing popularity in the tech sector.

A Controversial Exchange: An Invented Name and a Photo Never Sent

On X, user @prit4k reported that Instinct indicated a photo had "passed through" their conversation. The published screenshots show that the assistant allegedly provided the second name of a third party and mentioned an image from a document related to money, which the user claims they never sent. Instinct apologized and labeled the incident a "serious bug." The user then publicly questioned whether it was a data leak or a hallucination.

Noah Shinn's Defense and the Promise of an Anti-Hallucination Layer

Noah Shinn asserts that the incident is a result of a hallucination: the agent allegedly invented a proper name and amplified the error through its reasoning. He insists there was no data breach, no sharing of user information, and no isolation breach between accounts, without specifying how he determines that the reported case is a hallucination. He highlights existing protections, such as data partitioning, isolated sandboxes, short-lived local identifiers, and tool execution signed by identity. Shinn indicates that his team has been working for the past 48 hours on an active hallucination detection system, described as capable of anticipating or intercepting a forthcoming reasoning or tool call, with details to come in the following weeks. He did not immediately respond to a request for comment sent outside of business hours. Earlier this month, journalist Pranav Dixit tested the assistant and wrote that he was impressed by its intuitive help with daily tasks.