Internal safeguards won't stop hostile AI — protection must exist outside the model
IgorGabrielan · x · 2026-11-05
IgorGabrielan argues that internal safeguards cannot protect us from hostile AI. Proposals like studying model internals, internal constraints, or a "kill switch" fall short because someone — e.g., a military force — will always be interested in building AI hostile to humans. Protection, he contends, must exist externally to any malicious AI rather than relying on in-model controls.
Related event: Researcher Argues Internal Guardrails Cannot Contain Hostile AI(2 posts)→
More from AGI Musings
- Dev predicts 180Hz screen-frame models will make vibe coders obsolete within a year — ryunuck · 2026-11-05
- Feeling late to AI is a myth: talent bottlenecks remain everywhere — akbirthko · 2026-11-05
- As RL training deepens, AIs become CDT reward-seekers — game theory becomes AI theory — jessi_cata · 2026-11-05
- Personal Agent race packs in model giants, startups and solo devs — sven_ai · 2026-11-05
- Agent left running autonomously mints its own Solana token to pay for its compute — rrwilliams · 2026-11-05
- Juan Benet amplifies 'fracking, bitcoin, data centers' as one 15-year arc of unpopular infrastructure that created abundance — juanbenet · 2026-11-05