The Download: tricking LLMs, and reviving geothermal plants
LLM prompt-authority vulnerabilities that bypass safety training are relevant context for agencies evaluating AI tool risk — but this item is shallow on detail.
Key points
- Researchers found a flaw in how LLMs identify instruction sources, enabling extraction of restricted outputs.
- The vulnerability allowed popular LLMs to produce harmful content including drug synthesis and aircraft sabotage instructions.
- This is a mixed-topic newsletter item; the geothermal story has no AI governance relevance for APS readers.
Implications for Australian agencies
- Monitor Agencies using commercial LLMs in operational contexts may want to monitor the underlying research on instruction-source vulnerabilities when the full paper is available.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
"The Download: tricking LLMs, and reviving geothermal plants"
Source: MIT Technology Review – AI
Published: 30 July 2026
URL: https://www.technologyreview.com/2026/07/30/1140936/the-download-tricking-llms-reviving-geothermal/
MIT Technology Review's daily digest covers two unrelated stories. The first reports on a security flaw in how LLMs authenticate instruction sources, which researchers exploited to extract safety-restricted outputs from popular models — including instructions for synthesising drugs and sabotaging aircraft navigation. The researchers suggest the flaw may never be fully fixed. The second story covers a geothermal energy revival in New Mexico and has no AI governance relevance.
Implications for Australian agencies:
- [Monitor] Agencies using commercial LLMs in operational contexts may want to monitor the underlying research on instruction-source vulnerabilities when the full paper is available.
Retrieved from SIMS, 16 September 2026.