6. Stress Testing Deliberative Alignment for Anti-Scheming Training Builds a broad testbed for covert actions as a proxy for AI scheming, trains o3 and o4-mini with deliberative alignment, and shows big but incomplete drops in deceptive behavior.
SAFETY
-
GPT-4 Powered Malware MalTerminal Raises Security Concerns
By
–
GPT-4 Powered Malware Is Here: Researchers Call MalTerminal a Wake-Up Call https://
search.app/dRNQC #malware #GPT #LLMs #GenerativeAI #AI #ArtificialIntelligence #GenAI #security #CyberSecurity #CyberSec @lexfridman @KirkDBorne @Ronald_vanLoon @erikbryn @antgrasso @sallyeaves -
Code Hidden in LLM Layer Activations Demands Apology
By
–
The code was written at layers 22-30 and is stored in the value activations you just can’t read it. I think you owe the LLM an apology.
-
India AI Summit 2026: Safe AI for Social Good
By
–
As momentum builds towards the #IndiaAIImpactSummit2026 in New Delhi, this week’s pre-summit events spotlighted critical themes. 'Safe & Trusted AI for Social Good' in Hyderabad engaged experts on responsible and impactful AI adoption.
'Synergizing AI with Digital Humanism' -
ASI Alignment Requires Prior Political and Social Coordination
By
–
So ASI alignment is only interesting and important if we can *first* achieve political and social alignment of near-ASI technology. (If we navigate that successfully, then we'll actually be better placed to handle ASI as a global community anyway.)
-
AI Power Consolidation and Feudal System Risks
By
–
At this point, if it is used to consolidate power and create a permanent feudal system where 99.999% of humanity live in misery and can't do anything about it, then that's a problem.
-
Controlling Superintelligent AI Before Loss of Human Control
By
–
If there's some speculative future scenario where ASI is so powerful that humans can no longer control it, then there's an intermediate point before that where it's still extremely powerful, but humans still can control it.
-
AI Control Already Exists: Probability Assessment
By
–
People already have AI control right now. So the probability is 1.
-
LLM Security: Untrusted Text and Prompt Injection Risks
By
–
I'd add another circle to that diagram for the permissions of anyone else who authored text that made it into the context – which often means untrusted external attackers since it's so easy to sneak in malicious instructions for a lot of LLM systems
-
6 AI Security Guidelines for Healthcare Organizations
By
–
6 AI Security Guidelines for Healthcare Organizations
#AI #AIio #AIInnovation #ML #DataScience #Futureofwork @Scobleizer @AndrewYNg @drfeifei @KirkDBorne @fchollet @rowancheung @antgrasso