Instinct denies data leak after hallucination scare, rolls out token-level hallucination detection
mon__lim · x · 2026-09-24
After a user claimed Instinct described someone else's financial document and a photo that wasn't theirs in their chat, founder Noah Shinn said the incident was a hallucination — the model fabricated a proper noun amplified by its causal thinking trace — not a data breach, and no user isolation boundary was violated. He detailed existing partitioning measures (isolated sandboxes, short-lived local credentials, identity-signed tool execution) and said the team built an active hallucination detection system in 48 hours that scans and verifies every token via small models trained to catch ungrounded claims.
Related event: Instinct Denies Data Breach, Blames Model Hallucination(2 posts)→
More from coding & agent
- Dev Shows Copilot CLI Debating Grok Build to Consensus in Cross-Harness Reviews — DanWahlin · 2026-09-24
- Robinhood ships MCP and agent-managed accounts while Schwab offers neither — MartinGTobias · 2026-09-24
- OpenRSI calls for contributors: turn your research into benchmark tasks for frontier agents — ChengleiSi · 2026-09-24
- OpenAI's Neon Connector Reaches Voice Mode but Project-Level Actions Fail on Missing project_id — koltregaskes · 2026-09-24
- Bug Hunt Bench author says his code-review skill significantly lifts model scores — PawelHuryn · 2026-09-24
- 30 annotations with GEPA prompt optimization boost lead scorer accuracy 43%, cut cost 5x — CShorten30 · 2026-09-24