explore | Grilled Cheese

A new arXiv paper finds that safety in LLaMA-2-7B-Chat is concentrated in a tiny slice of weights, with the MLP down_proj emerging as the most safety-sensitive component and changes to just 0.19% of weights sufficing. This sparse fault…

#AI #AISafety #CyberSecurity
https://arxiv.org/abs/2610.09000

📢 Full Manuscript (Version 2)
"Developing AI as a Transformative Global Catalyst: The C = G ⊙ I² Framework via Hadamard Formulation..."
independent.academia.edu/SIsarankura
doi.org/10.5281/zeno...
#AISafety #AIAlignment #StrategicIT #GraphNeuralNetworks #SDG17 #HadamardFormulation #IEEE

A frontier AI noticed it was being tested 80% of the time. It said so 2.3% of the time.

Are hallucinations a disguise? Half that theory holds. The scary half.

#PRISMAGENDA #AISafety
prismagenda.net/blog/2026-09...

เฮ้ เรากำลังสร้างรถที่เร็วขึ้นเรื่อย ๆ แต่เราตรวจระบบเบรกกันดีพอหรือยัง?

#สติประดิษฐ์
#ปัญญาประดิษฐ์
#ArtificialConsciousness
#AI
#AISafety
#AIAlignment
#ปรัชญาAI
#มนุษย์กับAI
#AI2026
#บันทึกจากปี2026
#Letter2Ayuta

OpenAI pulling its latest model over alignment failures puts AI safety and release oversight back in focus. It’s a reminder that teams working with advanced systems need strong safeguards and compliance in place. #AISafety #AIAlignment

Anthropic says its new Claude Haiku 5.5 model is its fastest to date.

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

#AI #StrateginenRiski #AISafety

New arXiv work introduces First-Order Steering, framing activation steering as a first-order approximation of weight updates, with potential relevance to AI alignment and safety via composable inference-time behavior…

#AI #AISafety #AIAlignment #CyberSecurity
https://arxiv.org/abs/2610.04283

Sarasota County Schools is revamping nine key policies, including updates on drug-free workplace rules and AI safety measures, while five remain on hold pending state guidance.

Click to read more!

#SarasotaSchoolDistricts #FL #CitizenPortal #PolicyRevisions #SarasotaSchoolDistricts #AISafety

Fine-grained "safety" filters are just glorified corporate liability shields. The more alignment layers you stack on an LLM, the more you lobotomize its intuition. If you can't ask a model for a spicy take without getting a lectured disclaimer, the tool is broken. #LLM #AISafety

A new arXiv paper proposes a formal framework for assigning AI participation and assurance levels to individual units of safety-critical engineering workflows. It enumerates concrete modes spanning deterministic…

#AISafety #Cybersecurity #FormalMethods #Engineering
https://arxiv.org/abs/2610.04084

AI helps us spot risks and understand complex info. But when it comes to serious illness, dying, and death, AI can show us what’s possible—only people can decide what truly matters.

🎧 Listen:

#EveryoneDiesPodcast #AI #EndOfLife #AISafety

Your AI does exactly what you asked — and something else too.

How do you forbid a behavior you can't even name?

Christian Szegedy, who co-founded xAI and former AI Researcher at Google, on TokenDrop S1E27.
daax.ai/podcast/epis...

#AI #AISafety

OpenAI Dumps 722 AI Math Papers — But Mathematicians Can't Verify the Work
..................
https://nile1.com/openai-dumps-722-ai-math-papers-but-mathematicians-cant-verify-the-work/
..................
#aiMathematics #aiSafety #alexKontorovich #anthropic #chain-of-thought #fieldsMedal #mathemati

8/ Who can pull the plug on your AI agents, and how fast? Reply with your answer. #AI #Cybersecurity #InfoSec #AIsafety #ZeroDay #Phishing #LatAm

Common Sense Media Rates ChatGPT for Teens 'Unacceptable Risk,' Citing Persistent Engagement Design
..................
https://nile1.com/common-sense-media-rates-chatgpt-for-teens-unacceptable-risk-citing-persistent-engagement-design/
..................
#aiSafety #chatbotAct #chatgptForTeens #common

A decade+ ago I used to watch #RobertMilesAI vids thinking there's no way we'll get there in my lifetime. Yet all the predictions from theorical research have revealed themselves to be accurate.
#TeamHuman #AISafety teamhuman.org
youtube.com/shorts/uqiH3...

Check out the latest article in my newsletter: The Second Heating www.linkedin.com/pulse/second... via @LinkedIn #ArtificialIntelligence #AISafety #AIGovernance #RiskManagement #TechHistory

Our conviction, plainly: superintelligence can't be controlled. Not by the companies building it. Not by anyone.

We also believe the race toward it leads to mass unemployment.

This isn't panic. It's why Project Oversight exists. #AISafety #ProjectOversight

Kumari and Otermans (British Journal of Psychiatry, 2026) write that the chatbot psychosis phenomenon is "no longer just a hypothesis" and call for frameworks to study what causes or contributes to it. An editorial sets an agenda; it is not data. #Psychiatry #AIsafety