"LLMs are unsafe by definition": practitioners warn production agents fail silently without per-step checks

gerardsans · x · 2026-08-25

A debate on LLM safety argues LLMs are inherently unpredictable, fail silently, and accumulate errors, so stochastic components shouldn't run in production unless cleared at each step. A reply adds these issues have been documented in controlled studies since 2024 — Salesforce's AgentForce hit them late last year — but the industry never got the memo, prioritizing sales over safety. The commenter predicts the first major customer-facing agent product failure will be blamed on everything except the rogue agent itself.

Related event: AI safety debate: the real risk is bad engineering, not rogue models(5 posts)→

Original post →

More from coding & agent

coding & agent channel →