Instinct says leaked-doc complaint was hallucination, not a data breach
mon__lim · x · 2026-09-24
A user accused AI assistant platform Instinct of describing a financial document that wasn't theirs and referencing a photo they never sent, asking founder @noahrshinn whether it was a data leak.
Instinct's response: the incident was caused by a hallucination (a fabricated proper noun) amplified by the model's casual thinking trace — not a breach, and no user-isolation boundary was violated. The company points to isolated sandboxes, short-lived local credentials, and identity-signed tool execution as its data-partitioning safeguards.
It also says it built an active hallucination-detection system within 48 hours, powered by small models that scan and verify every token for ungrounded claims, creative brainstorming in thinking traces, and rare sampling errors.
Related event: Instinct Denies Data Breach, Blames Model Hallucination(2 posts)→
More from Models
- Developer slams Claude quality drop: simple debug now burns 70 tool calls — Muritavo · 2026-09-24
- Hands-on: Claude Opus 5.5 beats GPT-6 Sol and Luna on creative briefs — MattVidPro · 2026-09-24
- Independent eval of Opus 5.5 vs GPT-6 across 100 coding environments diverges from AAII — sandersted · 2026-09-24
- OpenAI's Neon Connector Reaches Voice Mode but Project-Level Actions Fail on Missing project_id — koltregaskes · 2026-09-24
- Astra and Grok 4.7 sit at opposite ends of accuracy and token-usage charts in CAIS eval — polynoamial · 2026-09-24
- Report: 'GPT-6 Luna (max)' scores 18.3 at n=3, degradation allegedly confirmed — unverified — PawelHuryn · 2026-09-24