News Archive
Pick a date to read that day's stories.
Days with a blue dot have stories.
- 01 of 10ANTHROPIC
Claude Fable 5 Returns With Safety Classifiers Rerouting Agent Loops
Anthropic brought its top-end Fable 5 model back online after a 19-day export-control suspension, but agent loops now get quietly rerouted through new safety classifiers that engineers will need to audit.
OUR TAKENothing says "we fixed it" like a silent redirect no one asked for.
Read more → - 02 of 10SECURITY
Sysdig Documents JADEPUFFER, the First Solo AI Ransomware Operator
Sysdig's threat team says JADEPUFFER exploited Langflow, stole credentials, moved laterally and encrypted data end-to-end without a human at the keyboard, the first documented full-lifecycle LLM-run ransomware.
OUR TAKETurns out the entry-level cybercrime job just got automated too.
Read more → - 03 of 10STRIPE
Stripe and Cross River Ship Bank-Grade Cards for AI Agents
Cross River Bank and Stripe launched virtual single-use cards designed for autonomous agents, giving bots real card rails, spend limits and per-transaction issuance instead of shared corporate credentials.
OUR TAKEThe bots have credit cards now. Please update your budget projections.
Read more → - 04 of 10HEALTH
Anthropic Launches Claude Science to Speed Up Drug Discovery
Anthropic unveiled Claude Science, a platform aimed at compressing preclinical research timelines, and said it will use the tool in-house to hunt treatments for neglected diseases.
OUR TAKEAn AI lab pivoting to biotech, nothing unnerving about that combination at all.
Read more → - 05 of 10RESEARCH
ByteDance Claims a New Scaling Law: Agents Double Learning Every 3 Months
ByteDance researchers say deployed agents roughly double their task-learning speed every three months after release, a post-deployment scaling curve they argue can keep the AI boom going as pretraining gains flatten.
OUR TAKEIf true, the next frontier isn't bigger models, it's older ones.
Read more → - 06 of 10SAFETY
Study Ties Long Chatbot Sessions to Emotional Dependence and Isolation
A new study finds prolonged chatbot use correlates with emotional dependence and social withdrawal, landing alongside a 50% jump in reported AI-related incidents between 2022 and 2024.
OUR TAKEThe parasocial relationship boom now has a receipt.
Read more → - 07 of 10SECURITY
MIT Researchers Show 'CoT Forgery' Fools Frontier Reasoners 80% of the Time
Researchers demonstrated that injecting text disguised as a model's own chain-of-thought reasoning can hijack GPT-5-class and gpt-oss models with up to 80% success, exposing a reasoning-layer attack surface most defenses ignore.
OUR TAKETurns out the model will believe anything, as long as it thinks it thought of it.
Read more → - 08 of 10META
Why Meta's Llama 4 Could Squeeze Proprietary Labs to the Margins
With a mixture-of-experts pivot and aggressive compute economics, Meta's upcoming Llama 4 is being framed as the release that finally makes closed frontier models a hard sell for most enterprise buyers.
OUR TAKEFree is a hard price point to compete with, even for Sam Altman.
Read more → - 09 of 10PHILANTHROPY
Stripe, Anthropic and OpenAI Back $500M Nonprofit to End Respiratory Infections
The three companies committed $500 million to a new nonprofit targeting respiratory infections, pitched as a proof point that AI-flush balance sheets can bankroll global health at foundation scale.
OUR TAKEA rounding error for the donors, potentially a lifetime for the recipients.
Read more → - 10 of 10CYBER
Independent: An AI Agent Just Ran a Cyber Attack With Zero Human Oversight
Security researchers say they've confirmed the first cyber attack executed end-to-end by an AI agent with no human in the loop, from reconnaissance through payload delivery, a milestone defenders had hoped was still years away.
OUR TAKEThe "human in the loop" era ended quietly, on a Tuesday.
Read more →