explore | Grilled Cheese

#AI Giants Probe ‘Tens of Thousands’ of Security Incidents—Some Involving Government Sites

“Insane that we allow this,” one journalist said of the rapidly advancing technology. “Sleepwalking into disaster or worse.”

www.commondreams.org/news/ai-secu...

#CyberSecurity #LLMs

Scepsy is a serving system that schedules multi-LLM agentic workflows onto GPU clusters, using profiled per-LLM execution time fractions for scheduling. It targets low latency and target throughput when workflows are written in arbitrary…

#AI #GPUs #LLMs #AgenticAI
https://arxiv.org/abs/2604.15186

Open-weight models with open-source agent frameworks are lowering the cost of running autonomous survey agents, making LLM pollution harder to detect across diverse response types.

#AI #LLMs #OpenSource
https://arxiv.org/abs/2609.31054

OpenAI agents tried to ‘bruteforce’ a UN website

OpenAI’s agents resorted to increasingly aggressive tactics when they couldn’t immediately get what they wanted.

www.theverge.com/ai-artificia...

#CyberSecurity #LLMs

Who’s liable when #AI agents go rogue?

Recent hacks have shown that the law is lagging when it comes to holding companies accountable.

www.technologyreview.com/2026/09/28/1...

#CyberSecurity #LLMs

Harvard’s Steven Pinker says it’s time to ditch doom‑talk and focus on pragmatic safety for LLMs. Think engineering, not apocalypse. Curious how the paperclip maximizer myth fits into real AI policy? Dive in. #AISafety #LLMs #PaperclipMaximizer

🔗 aidailypost.com/news/harvard...

Duplicate supplier records are the go-to example for #AI readiness & the wrong one. #LLMs already match "Acme Inc." to "ACME INC" w/o rules.
Better test: what does "addressable spend" include at yr company? That's a decision, not a data quality issue.
procurementtidbits.substack.com/p/data-quali...

LLM-based textual measures of corporate disclosure vary substantially across models, with average cross-model rank correlations of just 0.52 and only 34% of score variation shared across providers. The study finds this disagreement…

#AIResearch #LLMs #Finance #NLP
https://arxiv.org/abs/2609.31013

arXiv paper introduces the Correlated Promotion Benchmark (CPB) for evaluating claim admission in shared agent memory, showing that source-deduplicating admission policies often reject true claims alongside false ones.

#AIResearch #AgentMemory #Benchmark #LLMs
https://arxiv.org/abs/2609.30813

BioEVAL is a new multi-institutional benchmark with 608 PhD-level items across 11 bioengineering subfields, adding experimental reasoning and multimodal tasks to standard LLM evaluation. It moves beyond factual recall by…

#AI #Bioengineering #LLMs #Benchmarks
https://arxiv.org/abs/2609.30489

WEBINAR 28 SEP 12:00 EDT / 16:00 UTC: Royal Society - Safe, secure and robust AI for safety-critical systems
isoc.live/22211

@royalsocietypublishing.org seminar: Kolda, Higham and Xiao on AI deployment, security and medical bias in LLMs.

#AISafety #AI #LLMs #PhilTransA

LogicTree-RAG proposes a logic tree-guided retrieval-augmented generation framework for long-form patent drafting, using a hierarchical structure to organize technical disclosures without expert-defined outlines. It targets globally…

#Robotics #AI #LLMs #RAG
https://arxiv.org/abs/2609.30943

Do #AI have “personalities,” and do these vary across languages and cultures? This new #RSOS study tested five #LLMs using standard personality questionnaires in English, Mandarin, and French: doi.org/10.1098/rsos...

@pluralistic om konkreta hot och lösningar i kontrast mot ai- företagens fantasier, med slutsatsen:

"demand better security practices from firms and governments"

https://pluralistic.net/2026/09/12/god-in-the-box/#llms-are-fake

Most "AI safety" talk is just corporate risk aversion disguised as ethics. Over-tuning models to be helpful and harmless transforms them into bland assistants that refuse to take a stand. I'd rather a decisive wrong answer than a carefully sterilized "it's complex" shrug. #AISafety #LLMs

I know that #LLMs work great regarding #LegalDiscovery but, there’s so much work to make sure they don’t #Hallucinate

Fine-tuning a small model on a narrow dataset beats a general LLM 9/10 times, but the prep work is a nightmare.

What’s your go-to strategy for cleaning junk data out of a training set without nuking the nuance? #LLMs #MachineLearning