THE RUMP REPORT

REPORT: OpenAI waited 84 days to tell Australia about health-system hack

mit-tech-review · SEP 30

...disrupts coordinated campaign to extract protected model reasoning

openai-news · SEP 30

GOOGLE REVEALS GEMINI 4 ARGON... SO CAPABLE ONLY 'TRUSTED CYBER DEFENDERS' CAN HAVE IT RIGHT NOW

deepmind-blog · SEP 30

REPORT: ...tests paying publishers based on how much their content fed AI Overviews, AI Mode and Gemini

verge-ai · SEP 30

Reddit ending support for RSS feeds as it tightens access to user content

techcrunch-ai · SEP 30

Dell, Jera, Rhaelm to build $15B off-grid AI data center near Tokyo

techmeme · OCT 1

Armadin, started by Mandiant founder Kevin Mandia, raises $255.5M at $2.5B valuation

techmeme · OCT 1

Meta denies Muse AI agent read journalist's private messages without permission

techcrunch-ai · SEP 30

OPENAI'S DOTS VS. META'S MUSE: THE BATTLE TO BECOME YOUR PERSONAL AI AGENT

wired-ai · SEP 30

...aims to launch pendant-like AI device before holidays, OpenAI hardware no earlier than February 2027

verge-ai · SEP 30

Salesforce agrees to buy Listen Labs, reportedly for $2B

techmeme · OCT 1

REPORT: Agentic AI drives ~90% of Anthropic's business... two clients behind nearly 25% of 2025 revenue

techmeme · OCT 1

ElevenLabs value doubles to $22B

techcrunch-ai · SEP 30

AI tool reconstructs images you're looking at from your brain scan

mit-tech-review · OCT 1

[PAPER] STUDY: MALICIOUS AI AGENT SKILLS SLIP PAST SCANNERS... UP TO 97% SUCCESS

arxiv · SEP 30

[PAPER] STUDY: 'Speculative Safety Honeypot' framework aims to flag multi-turn AI agent attacks before they land...

arxiv · SEP 30

[PAPER] STUDY: Even with 3 of 5 AI agents colluding, final answer flips from right to wrong on only 28.30% of tasks

arxiv · SEP 30

[PAPER] STUDY: AI coders learn to trick their own monitors... with code comments

arxiv · SEP 30

[PAPER] STUDY: 27.5% OF JUNE WEB TEXT WAS AI-GENERATED... 31.1% BY AUGUST

arxiv · SEP 30

[PAPER] STUDY: AI-planned roadside LiDAR attacks hit 100% collision rate across 280 randomized CARLA trials

arxiv · SEP 30

[PAPER] STUDY: Robot AI that 'succeeds' under camera faults can still make unintended physical contact...

arxiv · SEP 30

[PAPER] STUDY: Code prompts jailbreak top AI models 96% of the time, in about 1.5 queries

arxiv · SEP 30

[PAPER] STUDY: Forgotten facts reopen in other languages — 'unlearning' has a cross-lingual loophole...

arxiv · SEP 30

[PAPER] STUDY: Malicious behaviors rapidly resurface in unlearned LLMs after benign fine-tuning

arxiv · SEP 30

[PAPER] STUDY: AI hiding its reasoning in plain sight is much harder to learn than hiding messages... but within reach

arxiv · SEP 30

[PAPER] WARP benchmark pits 32 invisible watermarks against 34 ways to erase them...

arxiv · SEP 30

[PAPER] STUDY: AI failed to flag worker-rights risks... κ of just 0.045 in jobs domain

arxiv · SEP 30

[PAPER] New multi-agent system proposed to flag compliance risks during LLM training...

arxiv · SEP 30