Debate: Is calling third-party model APIs a bigger security risk than weight backdoors?
jessi_cata · x · 2026-11-22
A debate on open-weight vs API-only model security. jessicata argues that when calling APIs from OpenAI, Anthropic or OpenRouter, not just the hardware owner but the software provider could be hacked — a risk absent when self-hosting known weights. mgaruccio counters that businesses already trust MS/AWS with all their data, so distrusting the same parties with prompts implies most leading security practitioners are wrong. jessicata clarifies the claim is a relative comparison of attack feasibility.
Related event: Debate: hacked API servers may pose more risk than weight backdoors(3 posts)→
More from Safety
- Sarcasm aimed at "zero data retention" claims — yenkel · 2026-11-22
- Honeypot catches Tencent crawler disguised as iPhone ignoring robots.txt to steal fake API key — rickasaurus · 2026-11-22
- Schneier and Frischmann: data minimization is imperative in the age of AI agents — bigdata · 2026-11-22
- After two weeks in China, author argues Western AI safety playbook doesn't transfer — and catching up is getting costlier — JacquesThibs · 2026-11-22
- US AI data centers' glass jaw: China controls ~two-thirds of the optical transceiver market — kevinnbass · 2026-11-22
- Eric Schmidt: a mutual AI race pause is impossible because it can't be verified — rohanpaul_ai · 2026-11-22