Today’s top line: The DOJ dismantled a large AI-driven student-aid fraud ring; OpenAI will honour a shared robots.txt signal to express AI training preferences.
DOJ uncovers AI-driven ‘ghost student’ scheme, seizes $750,000 crypto
Summary: The Department of Justice said investigators uncovered a large-scale scheme using AI to complete coursework and keep fake students enrolled, allowing fraudsters to siphon federal student aid. Authorities seized more than $750,000 in cryptocurrency from a network linked to Kenyan nationals, and the Department of Education identified over 830,000 potentially fraudulent federal student aid accounts, with total losses approaching $130 million if all are confirmed.
Digest: This shows AI is already being used at scale for financial crime that exploits public benefits systems; schools, lenders and fraud teams will need new detection tools and stronger identity controls. Watch for further arrests and tightened verification by education and financial authorities.
Source: Ars Technica
OpenAI will honour a shared robots.txt signal for AI training
Summary: OpenAI announced it will honour a shared robots.txt-style signal that lets website operators communicate their preferences about AI training to participating providers.
Digest: If adopted widely, the signal gives publishers a simple technical way to opt out of model training or signal restrictions, which could shift web crawling and data licensing behaviour across publishers and model builders.
Source: OpenAI
Trump declares foreign industrial robot imports a national security threat
Summary: A presidential proclamation issued October 28 brands imported industrial robots and related machinery as posing risks to US industrial competitiveness and national security. The directive asks the US and trading partners to adjust imports within 180 days and warns that tariffs, duties, quotas or fees could follow if agreements are not reached.
Digest: US manufacturers that rely on foreign-made robot arms and automation could face higher costs and supply disruption if trade measures follow; expect near-term uncertainty for procurement and increased pressure to localise heavy industrial-robot production.
Source: Ars Technica
Google’s prototype data-centre satellite is online, running Gemma 4
Summary: Google’s Suncatcher prototype satellite is live after launch earlier this month on a SpaceX Falcon 9. It carries four Google TPUs, generates up to 1 kW of power and can only run one TPU for roughly 20 minutes at a time; engineers commissioned subsystems over about three weeks and tested a DeepMind Gemma 4 model by asking it to write haiku and report orbital conditions.
Digest: Google is testing inference in orbit to study thermal and radiation effects on TPUs; if successful, follow-up satellites in 2027 will aim for higher power and inter-satellite optical links, signalling a long-term push to distribute inference to space hardware.
Source: TechCrunch
OpenAI tightens ChatGPT violence safeguards and adds teen protections
Summary: OpenAI said it is expanding violence-prevention measures in ChatGPT, introducing earlier interventions and committing to public safety reporting aimed at protecting teen users.
Digest: Platform-level changes like these aim to reduce harm and provide transparency; businesses and regulators will use them as reference points when assessing provider safety practices and compliance for younger users.
Source: OpenAI
DoorDash: third-party AI agents placed only ‘several hundred’ orders
Summary: DoorDash told investors it has seen only “several hundred” orders from third-party AI agents, versus its reported average of 3 million deliveries per day in 2025. The company said app-based orders still dominate while it explores agent integrations and potential advertising opportunities.
Digest: Agent-driven commerce remains nascent at scale; firms planning agent integrations should expect slow initial adoption and focus on service reliability and commercial models before assuming rapid volume shifts.
Source: TechCrunch
LRMoE: memory-aware MoE training cuts training time 21–45%
Summary: An arXiv paper introduces LRMoE, a memory-aware system for training sparse mixture-of-experts (MoE) models on memory-constrained nodes. By recalibrating backward demand, planning retention quotas and compacting communication, experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show cumulative MoE-layer training time reductions of 21.2% to 44.8% versus SYMI.
Digest: Training large MoE models on limited GPU memory is a common bottleneck; LRMoE’s techniques could lower cloud and on-prem compute costs and speed model iteration for teams using sparse-expert architectures.
Source: arXiv
Quick hits
- Fujifilm to quintuple LTO tape output for AI-driven archiving – Fujifilm plans a 5x production increase versus fiscal 2025 for its LTO Ultrium10 tape and is investing in its Kanagawa factory from fiscal 2026–2028 to supply cloud and data-centre archive demand (Ars Technica).
- Zest launches a text-based restaurant concierge that books tables – The startup’s AI concierge lives in SMS, can join group chats, uses Plaid-linked card data to learn tastes and books via OpenTable, Resy and SevenRooms (TechCrunch).
- ThinkJev paper offers budgeted ‘think-first’ decision method – ThinkJev runs a fast and slow decision path from one model so it only ‘thinks’ when needed; on JevBench Qwen3.5-27B met accuracy targets while reducing thinking tokens (arXiv).
- Report: Trump skipped intelligence about alleged Chinese spy near Nvidia CEO – New York Times sources say the president chose not to act amid concerns about harming Nvidia sales, a development that stalled some investigations (Ars Technica).