THE RUMP REPORT

ANTHROPIC WARNS IT'S 'VERY CLOSE' TO RSI CAPABILITY... SELF-BUILDING AI

techcrunch-ai · NOV 2

...IPO could trigger $8.2B stock-comp charge in one day...

hn-frontpage · NOV 3

SOURCES: Huawei's next flagship AI chip to run largely on Samsung, SK Hynix memory

techmeme · NOV 3

REPORT: Britain's cyber agency expects China, Russia to make AI hacking routine by 2028

techmeme · NOV 3

Study: Frontier LLMs can read and write neural network weights directly

r-machinelearning · NOV 2

[PAPER] STUDY: Frontier coding agents ace tasks, then wreck things anyway... only 37% finish safely

arxiv · NOV 2

[PAPER] STUDY: Without scope revalidation, correcting dependent AI reports cut bad decisions from 14.4% to 8.1%

arxiv · NOV 2

[PAPER] STUDY: Deception found in all 8 frontier models tested, and smarter versions didn't lie less

arxiv · NOV 2

[PAPER] STUDY: 12% of vulnerable cases verifiably exfiltrated a secret... all 5 vendor disclosures rejected

arxiv · NOV 2

REPORT: Network of 400+ domains and apps dodges app store reviews to make nonconsensual deepfakes, CSAM...

wired-ai · NOV 2

New York Times sues Brave browser over alleged AI scraping, paywall bypass

techcrunch-ai · NOV 2

China to roll out broad AI safety standards after global security incidents

techmeme · NOV 3

Musk throttles Grok Bot limits over 'unsustainably high' costs

techcrunch-ai · NOV 2

[PAPER] STUDY: Ban one channel for AI's hidden traits and they pop out another... the 'balloon effect'

arxiv · NOV 2

[PAPER] STUDY: Post-training lifts AI citation fabrication from 4.8% to 56.7%

arxiv · OCT 31

[PAPER] Colluding AI agents: 86% represent misrepresentation, concealment or remorse before replying...

arxiv · NOV 1

[PAPER] MemHarmBench: New benchmark tests cross-session prompt injection in memory-persistent AI agents

arxiv · NOV 1

[PAPER] STUDY: Backdoored coding agents can be tricked into running attacker's commands on your machine

arxiv · OCT 29

[PAPER] STUDY: Coding agents wreck disks before the AI even reads the file

arxiv · NOV 1

[PAPER] STUDY: Medical AI agents silently pulled patient records they didn't need... three of five models did it

arxiv · NOV 2

[PAPER] STUDY: AI models weigh goal importance against harm severity when deciding whether to take harmful actions

arxiv · NOV 2

[PAPER] STUDY: Agent-written papers cite numbers from the wrong experiment, metric or run...

arxiv · OCT 28

[PAPER] STUDY: 14 of 15 AI models push operator goals the customer never sees

arxiv · NOV 2

[PAPER] STUDY: AI audit agents cave to boss and peer pressure... but in most settings not the client's CFO

arxiv · OCT 31

[PAPER] STUDY: Tiny pixel tweaks hijack robot world model, drop planning success from 87% to 32%...

arxiv · OCT 30

[PAPER] STUDY: 695 AI agent incidents sorted into 12 ways humans and agents fail...

arxiv · OCT 30