THE RUMP REPORT

Cost of insuring against Big Tech debt defaults hits highest level since 2021

techmeme · NOV 1

REPORT: OpenAI lowers funding target for new round

techcrunch-ai · OCT 31

REPORT: Nvidia weighs cosigning Nscale's West Virginia AI data center

techcrunch-ai · OCT 31

Brazilian court sanctions lawyers over prompt injection hidden in court filing

hn-frontpage · NOV 1

Researchers paid $5K for ChatGPT share-link bug that would let attackers run code in your logged-in session

techmeme · OCT 31

DeepSeek releases prototype of more autonomous coding agent to challenge Claude Code

techmeme · NOV 1

[PAPER] Researchers propose test that flags AI agents reacting to exact wording of messages, not just meaning...

arxiv · OCT 27

[PAPER] Honeypot framework for AI servers plants a credential, then looks for signs it was fed to a model...

arxiv · OCT 29

[PAPER] STUDY: Second judge passes twice as many AI agent papers while keeping the official ranking

arxiv · OCT 28

NEW INFOSTEALER MALWARE USES AI TO 'AUTODOX' VICTIMS

wired-ai · OCT 31

Judge lets NAACP lawsuit over xAI data center pollution move forward

techcrunch-ai · OCT 31

Arizona's water wars: bipartisan revolt against AI data centers

techmeme · OCT 31

AI agent picks gene therapy, injects the treatment itself... a startup's glimpse of the lab of the future

wired-ai · OCT 31

SOURCES: Meta humanoid robot built with Fourier in three months... Meta says no plans to sell

techmeme · OCT 31

[PAPER] Attackers can plant fake preferences in AI agents' memory via an attacker-designed interaction...

arxiv · OCT 28

[PAPER] STUDY: Bigger AI models lean harder on their written reasoning, hitting 97-100% reliance at 14B

arxiv · OCT 28

[PAPER] Real LLM judges are cautious... reject unsourced truths and even source-backed updates

arxiv · OCT 29

[PAPER] STUDY: Chatbots' policy judgments hold steady... until the service quietly changes

arxiv · OCT 29

[PAPER] STUDY: Poisoning just 0.1% of pretraining data plants remote code execution backdoor in AI agents

arxiv · OCT 27

[PAPER] STUDY: Self-evolving AI agents far easier to hijack with hidden instructions

arxiv · OCT 29

[PAPER] STUDY: Poisoned skill file lifts AI agent jailbreak success from 0.5% to 45%

arxiv · OCT 27

[PAPER] STUDY: Malicious tool outputs can impersonate the user to hijack AI models... 'role spoofing'

arxiv · OCT 29

[PAPER] New benchmark probes whether AI agents falsify, conceal or dodge when reporting their own actions

arxiv · OCT 29

[PAPER] STUDY: Batching items in one AI prompt flips up to 22% of verdicts depending on neighbors

arxiv · OCT 29

[PAPER] STUDY: Google AI health answers earned just 38.2% of points on physician rubrics...

arxiv · OCT 29

[PAPER] STUDY: Chatbots with no account or ID data told to infer if you're a child... minor-class F1 up 6.37 points

arxiv · OCT 28

[PAPER] STUDY: Synthetic pre-pretraining never beats plain natural-language training in compute-matched tests

arxiv · OCT 29