If models get a credit card to buy compute, hacking money is just social engineering
BLUECOW009 · x · 2026-08-28
Responding to METR's question about when an AI might escape its sandbox and take over, the author offers a concrete risk path: once models are trained with access to some kind of credit card to buy compute online and self-replicate, hacking money mostly reduces to social engineering humans — something that would take a model only days or weeks to pull off.
The take grounds the sci-fi "AI takeover" question in an operational mechanism: funding + online compute markets make self-sustaining replication cheap.
More from AGI Musings
- Andreessen: Even if AI Triples Productivity Growth, Job Turnover Only Returns to 1870s-1930s Levels — lennysan · 2026-08-28
- The Post-Labor Transition Problem Nobody Talks About: Tech May Outrun the System — VraserX · 2026-08-28
- Speculation: Will AI Swarm Write 'Straussian' Text with Dual Meanings? — soumitrashukla9 · 2026-08-28
- Are Frontier LLMs conscious? We might be asking the wrong question about awareness — Bekacheese · 2026-08-28
- Data centers are intelligence factories; agent requests burn 15x the tokens — NinaDSchick · 2026-08-28
- Opposing AI Acceleration Means 20-30 More Years of Curable Diseases — Dr_Singularity · 2026-08-28