PHP Patches Two Critical RCEs That AI-Assisted Researchers Turned Into Working Exploits -- After a Root Cause First Spotted in 2024 Stalled, and an Embargo That Leaked a Day Early
PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two memory-safety flaws in how the engine survives out-of-memory and fatal errors. Calif, OpenAI and Wiz all reported them with help from frontier models. The PHP Foundation's own account is the more useful read: the bug was not new, the fix was hard, and the disclosure plan broke under a leak.
LiteLLM 1.107.0 Shipped a Credential Stealer to PyPI and Docker for About Two Hours -- and LiteLLM Believes the Stolen GitHub Tokens Came From Its Own Proxy's Logs
The popular open-source LLM gateway published a poisoned release built from a lookalike repository. The payload ran on every Python start, went after provider keys stored in the proxy database, and stolen AWS credentials were checked for Bedrock access within hours. LiteLLM's account of how the attacker got in is still provisional.
Google Cloud Will Sell OpenAI's Models and Plans Up to 960,000 Rubin GPUs, While NVIDIA Says 'Capacity Is the Constraint' and Pays Suppliers to Add It
On October 21 OpenAI said its models and Codex are coming to Google Cloud's Model Garden, and NVIDIA's GTC Berlin brought Vera Rubin A5X instances, confidential computing on bare metal for model builders, and funded expansion of substrates, packaging and testing. Most of it is announced rather than available.
Zomato's Parent Eternal Unveils Dark, a Private AI Chat Built on Fine-Tuned Open-Weight Models -- and Names Its Events Business as Most Exposed to AI Agents
Eternal's Q2 FY27 shareholder letter pitches Dark at founders, researchers and governments on the promise that 'We cannot read your conversations, even if we are forced to.' The same letter grades which of its own businesses AI agents could bypass. Neither claim has been tested outside the company.
Google's Blind-Sweep Ultrasound Model Matches Expert Sonographers in a Prospective Two-Country Study -- Including in Nairobi, Where It Saw No Local Training Data
Novice operators glided a low-cost handheld probe across the abdomen, and a model running on a phone estimated gestational age with a mean error of 4.18 days, against 4.48 for expert biometry. The study is notable for being prospective, live and on unfamiliar hardware.
Framework Watch: Claude Code 2.1.307 Emits an Event Every Time a Guardrail Fires, Pydantic AI 2.59 Stops Passing Background Tool Results Off as User Text, and ADK Java 1.13 Blocks DNS Rebinding
Releases and commits from October 21 and early October 22 UTC that earlier Framework Watch entries did not cover. The common thread is that the harness now labels its own interventions and treats the data it handles as untrusted. Each item comes from the project's release notes or commit history.
Research Radar: Stored Lessons Make Agents Repeat Their Mistakes, a Decision-Model Controller Adds 19% Cost for No Gain, and Agents That Edit Their Own KV Cache Cut Cost 27.7%
Three arXiv preprints from October 17-20 listed on Hugging Face's October 21 and 22 Daily Papers. Each tests a component teams are adding to agent stacks right now: persistent memory, cheap controller models and context compaction. All results are the authors' own and not yet peer reviewed.
Trend Watch: The Assistant Moves Off Your Screen -- Claude Duet and Notion's Rebuilt Agents Each Get Their Own Computer, and the Design Problem Becomes When to Interrupt You
In three days, Anthropic launched a proactive agent that works from its own virtual machine and messages you first, and Notion said its agents now have their own computer and browser and can run for up to 24 hours. The interesting choices are not about capability: they are about check-in frequency, approval gates and who holds the credentials.