ChatGPT links financial accounts as Claude Mythos aids M5 exploit

OpenAI introduced a personal finance feature in ChatGPT on May 15, 2026. It is rolling out gradually to Pro users in the United States on the web and iOS. Through Finances, users can connect accounts via Plaid and ask ChatGPT questions grounded in their own financial data. The feature is part of ChatGPT rather than a separate app.¹²

Once accounts are linked, ChatGPT syncs and categorizes the data for analysis.¹ Reported access is read-only: ChatGPT cannot move money or see full account numbers.³⁴ The feature takes OpenAI beyond general financial questions into analysis based on a user's account information. That makes the decision to link sensitive data an important consideration. For now, availability remains limited to US Pro users.¹

Security research team Calif reported on May 14 that it had used Claude Mythos Preview to help develop a macOS kernel exploit for Apple M5 hardware.¹² The attack chain gave a low-privilege local user root access while Memory Integrity Enforcement (MIE) was enabled.¹² Calif described the model’s role in development; the disclosure did not come from Anthropic.¹

The case shows how an AI model can assist with a complex exploit: Calif says it took about five days to go from finding the vulnerabilities to a working attack.¹² This was not a remote attack. An attacker would first need to run code as a local user.¹² A successful test with MIE enabled also does not establish that the protection is ineffective on every device.¹²

Nous Research introduced Token Superposition Training (TST) on May 13, 2026, as a change to the standard LLM pretraining loop.¹ Early in training, the method groups contiguous tokens and averages their embeddings. It later returns to ordinary next-token prediction.¹ The company says TST delivers a two- to threefold reduction in wall-clock training time at matched compute (FLOPs).¹

The method was tested on several dense models and a mixture-of-experts model.¹ For the latter, a write-up reports a lower final training loss and fewer B200 GPU-hours than a baseline run at matched FLOPs.² The results suggest shorter training runs for a given compute budget, but they do not establish that implementation takes no additional work. Whether the gains hold across other architectures, data mixtures, and training setups remains unclear.²¹

On May 14, OpenAI introduced Codex in the ChatGPT app for iOS and Android as a preview.¹ Users can review results, start tasks, and steer ongoing sessions from their phones.¹ Coding work still runs on a connected computer, not on the phone.¹² The feature is rolling out across all ChatGPT plans in supported regions.¹

Codex was already available as a coding tool; the update adds mobile access to existing work environments.¹² The ChatGPT app is not a standalone code editor for phones.¹² Mobile connectivity initially supports macOS, with Windows support planned for later.¹² For teams, the change offers another way to monitor and direct ongoing tasks away from a desk, rather than a new environment for running code.¹²

xAI introduced Grok Build in mid-May as an early beta coding agent for the terminal. The tool is intended to help with software development and multistep tasks. Access was initially limited to SuperGrok Heavy, a $300-per-month tier.¹² In a May 25 announcement, xAI also listed SuperGrok and X Premium Plus subscribers as eligible for access.³

Grok Build enters a market where developers can already use terminal-based coding agents such as Claude Code.¹² For teams, the question is not just whether they can get access, but how reliably the tool works in their own repositories. xAI still describes Grok Build as an early beta and says it plans to improve it using user feedback.³ The announcement does not establish a reliable performance comparison with competing tools.

More than 50 researchers and engineers have reportedly left xAI since SpaceX acquired it in February 2026.¹² TechCrunch described a mix of layoffs, firings, and voluntary departures.¹ Those who left reportedly include leaders involved in coding features, world models, and Grok voice features.¹ The companies have not publicly confirmed the departure count.¹

Losing experienced staff could make it harder to develop new models and features, but departures alone do not establish how future products will perform. The team responsible for pre-training new models has reportedly shrunk to a handful of people; that figure also lacks official confirmation.¹ Former xAI employees have moved to Meta and Thinking Machines Lab, among other destinations.¹² The effect on research capacity and product plans remains unclear.

X published code for its For You feed recommendation algorithm on GitHub on May 15. Elon Musk announced the release that day.¹ The published ranking code uses a Grok-based transformer model called Phoenix to score posts.² The release gives developers a way to examine part of the technical basis for recommendations in a major public feed.

The view is limited: the release does not establish that the production Phoenix model weights or every component of the live feed are public.³² Available code and a smaller Phoenix model support testing, but do not guarantee a reproduction of the rankings users see on X.³⁴ Assessments of the system therefore need to distinguish the published code from its production configuration.

On May 16, 2026, Unsloth updated the instructions for its Qwen3.5-9B-MTP-GGUF files.¹ The files enable speculative decoding with multi-token prediction (MTP) when the runtime supports it.²¹ Unsloth says generation can be roughly 1.5 to 2 times faster. That is the provider’s claim, not a guaranteed result for every deployment.²

This is not a new Qwen base model; it is an MTP variant for an existing model.²¹ It may interest teams running Qwen locally that want faster output without changing base models. Results depend on the setup: one user reported slower output with MTP in a discussion of the model.³ Teams should therefore measure performance on their own hardware and workloads before changing their inference configuration.

Moonshot AI introduced Kimi WebBridge on May 14, 2026. The tool combines a browser extension with a local background service. It is designed to let coding agents control existing Chrome and Edge sessions through the Chrome DevTools Protocol. That gives agents access to the browser a person already uses, rather than requiring a separate session for web tasks.¹²

Access to signed-in sessions may make website tasks easier, but it also makes control of those sessions important. Kimi says logins, cookies, and page content remain on the user’s device. The cited sources do not provide an independent technical verification of that privacy claim.¹² Companies must also account for the fact that an agent with access to an active session can take actions as the signed-in user.²