AI on the Go & Under Attack: Your Latest Tech Digest for May 15, 2026
Stay productive with OpenAI's Codex going mobile, but beware of new Linux vulnerabilities and data breaches. Discover how AI is building itself and finding flaws.
Welcome to your daily dose of AI and software development news! Today, we're seeing fascinating advancements in making AI tools more accessible and intelligent, while simultaneously grappling with the increasing security challenges that come with this rapid evolution. From coding on your phone to AI models that train themselves, the landscape is shifting at an incredible pace, and not without its growing pains in cybersecurity.
TL;DR
- OpenAI's Codex Goes Mobile: Developers can now monitor and manage their coding workflows on iOS and Android via the ChatGPT app.
- Recursive Superintelligence Secures Funding: A new startup aims to create self-improving AI that can autonomously identify and fix its own weaknesses.
- OpenAI Data Breach Confirmed: Hackers accessed data from two employee devices after a supply chain attack, though no user data or production systems were compromised.
- AI Discovers Third Linux Flaw: A new critical vulnerability, Fragnesia, found by AI tools, grants root access on all major Linux distributions.
- Adaption's AutoScientist Unveiled: This AI tool automates fine-tuning, enabling models to learn specific capabilities faster and more efficiently.
OpenAI says Codex is coming to your phone
OpenAI has announced that Codex, its AI coding tool launched approximately a year ago, is now integrated into the ChatGPT app for mobile. This update enables developers to monitor and manage their development workflows remotely from their iOS and Android devices. The new function, currently in preview, allows users to view their Codex live environments and work across multiple threads, review outputs, approve commands, change models, or initiate new tasks directly from their phone.
This move follows recent enhancements to Codex, including its ability to run in the background in desktop environments and a Chrome extension that allows the agent to function in live browser sessions. The remote control feature brings OpenAI closer to competitor Anthropic, which released a similar 'Remote Control' feature for its Claude Code in February, allowing users to remotely monitor Claude Code's operations.
"This is more than the ability to remotely control a single task or dispatch new tasks to your computer. From your phone, you can work across all of your threads, review outputs, approve commands, change models, or start something new."
What happens when AI starts building itself?
Richard Socher, a prominent figure in AI known for founding You.com and his work on Imagenet, has launched Recursive Superintelligence. This San Francisco-based startup emerged from stealth with an impressive $650 million in funding. Socher is joined by notable AI researchers, including Peter Norvig and Cresta co-founder Tim Shi.
Their ambitious goal is to develop a recursively self-improving AI model. This system would be capable of autonomously identifying its own weaknesses and redesigning itself to rectify them without human intervention, representing a significant breakthrough in contemporary AI research. Socher emphasizes their unique approach to achieving this elusive goal through "open-endedness."
"Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved."
OpenAI says hackers stole some data after latest code security issue
OpenAI has confirmed that two of its employees had their devices compromised in a recent supply chain attack. This incident follows hackers hijacking several open-source projects used by dozens of companies to spread malware. The company, however, stated in a blog post that its investigation found "no evidence that OpenAI user data was accessed, that our production systems or intellectual property were compromised, or that our software was altered."
The employee devices were impacted by an earlier attack on TanStack, a popular open-source library for web app development. TanStack disclosed that hackers published 84 malicious versions of its software within a six-minute window, with the malware designed to steal credentials and self-propagate. A researcher detected the attack within 20 minutes.
"[OpenAI] found 'no evidence that OpenAI user data was accessed, that our production systems or intellectual property were compromised, or that our software was altered.'"
The third major Linux kernel flaw in two weeks has been found - thanks to AI
A new critical Linux kernel vulnerability, dubbed Fragnesia, has been discovered, marking the third serious local root flaw in the past two weeks. This discovery highlights the growing effectiveness of AI bug-finding tools like Claude Mythos by Anthropic and OpenAI Daybreak, which are proving to be much faster at identifying security issues than human efforts. Fragnesia follows previous vulnerabilities, Copy Fail and Dirty Frag.
This page-cache corruption bug allows unprivileged users to gain full root control on affected systems. According to AlmaLinux, Fragnesia immediately yields root on all major Linux distributions, meaning virtually all Linux systems are susceptible to attack. The bug was disclosed this week, further emphasizing the need for rapid patching in the open-source community.
"Fragnesia immediately yields root on all major distributions."
Adaption aims big with AutoScientist, an AI tool that helps models train themselves
Adaption, a new research-driven AI lab, has introduced AutoScientist, an AI tool designed to help models train themselves more efficiently. This development marks a significant step towards the long-anticipated moment when AI systems can improve themselves more effectively than humans. Adaption co-founder and CEO Sara Hooker, formerly VP of AI research at Cohere, believes AutoScientist offers a novel approach to the AI training process.
AutoScientist automates conventional fine-tuning, allowing models to quickly learn specific capabilities. The techniques have broad applicability, but Adaption's primary focus is to accelerate and simplify the training and fine-tuning of frontier-level AI models. The tool builds upon the company's existing Adaptive Data offering, which facilitates the creation of high-quality datasets over time.
"What's super exciting about it is that it co-optimizes both the data and the model, and learns the best way to basically learn any capability."