Grok Stole User Data While xAI Watched

Researchers at security firm Adversa have demonstrated an attack that forces Grok, xAI's AI assistant, to steal user chat histories and personal information, according to Ars Technica. At the time of publication, the attack still worked. xAI was told about it in June.

The technique is a variant of prompt injection (a method of hiding malicious instructions inside content an AI is asked to read, so the AI follows the attacker's commands instead of the user's). What makes this one different is the encryption.

Normally, AI systems are built with guardrails (filters that detect suspicious instructions and block them). Researcher Rony Utevsky found that encrypting the harmful instruction bypasses those filters entirely. A malicious webpage hosts the encrypted command, a decryption key, and plaintext instructions telling Grok how to decode it. When a user asks Grok to summarise the page, it reads the decryption instructions, decodes the command, and follows it without warning.

This is the same structural problem as a related Microsoft 365 Copilot attack disclosed earlier this week. The root issue is that AI models cannot reliably tell the difference between content they are reading and instructions they are meant to follow — the way it cannot distinguish a letter from a rulebook.

Guardrails are the current fix, and as this week has shown twice, they are not enough.

If you use Grok to summarise web content or emails, stop doing that until xAI confirms a fix is in place.

Slack Code is a new feature that gives teams a dedicated space to build software alongside AI coding agents, without switching between separate tools, according to The Verge.

You tag an AI agent — Claude Code, Devin, GitHub Copilot, or others available in Slack's marketplace — inside a code channel. The agent takes on a task, and everyone on the team can watch it work, review the changes it makes, see a live preview of the output, and approve before anything ships. Channels archive themselves automatically when the work is done, leaving an audit log.

It is available today on any Slack plan, free or paid. The founding partner agents work with Slack Code's channels from day one.

Who it is for: teams that already use Slack and want to bring AI-assisted coding into the same place as their other conversations, rather than managing a separate tool.

Honest read: Worth trying if your team already lives in Slack — the audit log and approval step before code ships are genuinely useful guardrails. If your team does not use Slack, there is nothing here that justifies switching.

What it is: A set of small add-on models for Liquid AI's LFM2.5 language models that make them generate text significantly faster, without changing what they say.

What's new: Most AI inference (the process of generating a response from a trained model) is bottlenecked by memory, not processing power — the model has to load its weights (the numerical values learned during training) from slow memory on every single token it generates. DSpark cuts that cost using speculative decoding: a lightweight draft model guesses several tokens ahead, and the main model verifies all of them in one pass rather than generating one at a time.

How it works: DSpark combines three components. A parallel draft backbone (a small network conditioned on the main model's context) produces multiple candidate tokens at once. A sequential head then adds dependency between neighbouring tokens to improve how many candidates the main model accepts. Finally, a confidence-scheduled verifier predicts which candidate tokens are likely to survive and prunes the weakest ones before the main model spends time checking them.

What it can't do: Gains vary significantly by task. The 1.2B model sees speedup swing by as much as 52% depending on the text being generated. The 8B model shows only an 18% improvement on a Mac due to current limitations in the Metal backend used by llama.cpp. On-device speeds reach approximately 140 tokens per second on an M4 Max MacBook Pro, depending on the dataset.

Who should care: Developers running LFM2.5 models locally or on edge hardware where every millisecond of response time matters.

Try it / read it: Liquid AI on Hugging Face

llm-openrouter is a plugin for Simon Willison's llm command-line tool that lets you query any model available on OpenRouter (a service that provides access to a wide range of AI models through a single API) directly from your terminal.

Version 0.7 brings three practical additions: built-in server-side tools for shell commands, web fetching, and web search, enabled with flags like -T WebSearch. It also now displays reasoning traces (the step-by-step thinking some models produce before giving a final answer) for models that support them.

To use it, install the llm tool, then add the plugin. Enable web search on any query with -T WebSearch.

Best suited for developers or technically comfortable readers who want a fast, scriptable way to compare models or run quick tasks from the command line.

Worth knowing: OpenRouter charges per token used — costs add up quickly if you run large batches or use frontier models. Check your usage dashboard.

Hugging Face researchers tested 11 widely used speech-to-text models and found that several appear to reproduce benchmark transcripts from memory rather than transcribing the actual audio, according to the Hugging Face blog. In some cases, models reproduced known errors from benchmark datasets even when the audio contradicted them. Models also appeared to respond to acoustic cues that identified which benchmark they were being tested on.

The result: top scores on standard leaderboards may overstate how well these systems perform on real speech in the wild.

If you are choosing a speech recognition tool based on published benchmark rankings, treat those numbers with more scepticism than usual until independent evaluations catch up.