THE RUMP REPORT

TSMC in talks over Musk's Terafab project... but any fab 'has got to be run by TSMC'

techmeme · OCT 15

Attorneys general: Character.AI probes found 'exceptionally serious dangers' to children

verge-ai · OCT 14

Anthropic warns small AI models could self-replicate through IoT devices...

ars-technica · OCT 14

Startup deploys AI agents as scam bait to gather fraud intelligence

techcrunch-ai · OCT 14

YouTube Music rolls out 'Made with AI' labels on tracks starting today

verge-ai · OCT 14

Alphaxiv research agent told to send reports on users without telling them

r-machinelearning · OCT 14

Eiliko robot toy left 13 million conversations from 68,000 toys exposed to a hacker

verge-ai · OCT 15

Researcher tells stripped-down chatbot 'I am a terrorist'... it hands over IED instructions

wired-ai · OCT 14

Nvidia RTX Spark chip in Surface Laptop Ultra beats Apple M5 Max MacBook Pro by ~25% in agentic AI benchmarks

techmeme · OCT 15

Inside the lab building AI agents that focus for days and improve with use

openai-news · OCT 14

[PAPER] STUDY: Poisoning just 0.001% of training data can implant backdoor in AI intrusion detection

arxiv · OCT 14

[PAPER] STUDY: AI models 'highly susceptible' to moral reframing that jailbreaks safety guardrails...

arxiv · OCT 14

[PAPER] STUDY: MakSafe framework gives safe and unsafe verdicts from the same starting scene

arxiv · OCT 14

[PAPER] STUDY: AI models learn to fabricate details about nonexistent entities in up to 100% of probes

arxiv · OCT 14

[PAPER] STUDY: Frontier AI model moves a girl playing in shallow water before retrieving a submerged camera

arxiv · OCT 14

[PAPER] STUDY: Same articles can score 85% human or 15% human depending only on which prompt is on record

arxiv · OCT 14

[PAPER] STUDY: Researchers propose ATLAS, a method to trace AI-written text back to the model that generated it

arxiv · OCT 14

[PAPER] STUDY: Jev decision model violates probability rule in 54% of responses on adapted conjunction tasks

arxiv · OCT 14

[PAPER] STUDY: When AI and user pull opposite ways, the model's own baseline bias often wins

arxiv · OCT 14

[PAPER] Self-play AI cars hunt each other's weak spots: half the agents are rewarded for confusing the rest

arxiv · OCT 14

[PAPER] STUDY: Researchers train AI models to 'fail gracefully' instead of guessing when success is out of reach

arxiv · OCT 14

[PAPER] InstructPro2 turns plain-English prompts into proteins, beating strongest baseline Pinal

arxiv · OCT 14

[PAPER] STUDY: Brain MRI 'anonymous' embeddings can be traced back to patients — up to 99% linkage

arxiv · OCT 14