In partnership with |
 |
|
|
|
| |
| |
Good morning, AI enthusiast. |
xAI just launched Grok Bot, a beta of AI "teammates" that sign into your actual tools - Gmail, your CRM, internal dashboards - and grind through work on a cloud computer while you're offline. Built with Cursor, the bots message you like a colleague instead of a chatbot, then report back once the job's done. |
It's xAI's clearest move yet from "assistant you prompt" to "coworker you delegate to," landing the same week researchers found a way to peel open supposedly private AI reasoning and Anthropic quietly started fingerprinting everything Claude writes. If your AI is working while you sleep, reasoning where you can't see it, and signing its own outputs, how much of that do you actually get to watch? |
Today in AI Brief: |
xAI's Grok Bot works your tools 24/7 A crack in AI's "encrypted" reasoning Claude starts invisibly watermarking everything you generate |
|
|
|
| |
|
10x the context. Half the time. |
|
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone. |
Try Wispr Flow free |
| |
| |
xAI's Grok Bot Is an Always-On AI Teammate |
In Brief: xAI launched Grok Bot, a beta of AI agents that work autonomously on cloud computers around the clock, signing into a user's real tools and reporting back once a task is finished - built in partnership with Cursor. |
The Details: |
Bots sign into a user's actual CRM, Gmail, product UI, and internal enterprise tools without needing clean APIs or MCP support, so they operate the way a human employee would. Multiple bots can coordinate in parallel, messaging each other and handing off work, with early use cases spanning sales outbound, CRM upkeep, hiring ops, expense processing, and bug triage. Access is limited during this early beta to SuperGrok Heavy and Cursor Ultra/Teams Premium subscribers on macOS and iOS, with an enterprise waitlist open for broader rollout. |
Take Away: |
Grok Bot pushes xAI into the agentic-teammate race already crowded by rivals' coding and computer-use agents - the pitch isn't a smarter chatbot, it's a coworker that clocks in without being asked. Whichever lab nails "acts like an employee" first captures the workflows that actually get automated. |
|
|
|
| |
|
| |
| |
A New Attack Cracks AI's "Encrypted" Reasoning |
In Brief: A new paper, Stealing Reasoning Traces from Proprietary LLM APIs, found that the encrypted reasoning blocks returned by OpenAI, Anthropic, and Google are interchangeable across sessions, users, and even models - feed one into a weaker sibling model with a jailbreak prompt, and it decrypts back into plain text without ever touching the actual encryption key. |
The Details: |
Scraping 315,320 public reasoning blocks, the researchers recovered 367 pieces of personal information and 182 credentials - API keys and passwords users assumed were sealed inside "private" reasoning. The same trick exposes hazardous content models scrub from their final answers but leave sitting in the reasoning trace, and lets attackers hide invisible prompt injections inside public agent deployments. The team found evidence consistent with model distillation - a rival training on leaked reasoning - though similarity alone doesn't prove it; all three labs were notified before publication and have since patched their systems. |
Take Away: |
Encrypting reasoning was supposed to protect both users' hidden inputs and labs' competitive edge; instead it created an attack surface nobody was watching. Expect every major lab to start treating reasoning traces with the same security scrutiny as the model weights themselves. |
|
|
|
| |
|
|
[Webinar] Can you prove AI is working? |
|
AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem. |
Join live on Aug 19 (FREE) to see: |
The 4 metrics to measure the gap where gains leak out before production. The 8 stages of context maturity, the specific walls capping your metrics, and a free tool to pinpoint where your team is Why more MCPs and bigger context windows aren't enough and what it takes to get real value from your agents. |
Register Now |
| |
| |
Claude Now Invisibly Watermarks Everything You Generate |
In Brief: Anthropic quietly began embedding invisible watermarks into everything Claude generates - text, code, and files - across the Claude Platform, Claude Code, Claude Cowork, and Claude Tag, plus cloud partners AWS, Google Cloud, and Microsoft Foundry. |
The Details: |
Models launched on or after August 2 ship with watermarking by default; older models are being retrofitted, and Anthropic plans to release its own detection tool. Text watermarks survive copy-pasting elsewhere, though heavy editing can strip them, while generated files like .svg, .png, and .jpg get C2PA provenance metadata that flags tampering. Anthropic frames a detected watermark as meaning content "may have been processed" by Claude, not that Claude authored it - a hedge that matters since U.S. copyright law requires human authorship for protection. |
Take Away: |
After years of fights over whose data trained AI models, Anthropic is now the one stamping its own fingerprint onto everything that comes out, whether users asked for it or not. With xAI notably skipping similar measures, watermarking is shaping up as another front in how labs differentiate on trust versus openness. |
|
|
|
| |
|
| |
| |
Everything else in AI |
River AI raised $1.1 billion for an open-source AI startup founded by ex-xAI co-founder Igor Babuschkin, who says he doesn't want "these AI companies to rule the world." |
NVIDIA partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to launch financing platforms aiming to mobilize more than $500 billion for AI compute infrastructure. |
Anthropic struck a reported $9.1 billion compute deal with Riot Platforms, tapping 191 megawatts from the crypto miner's former mining operation. |
Lightricks released LTX-2.5, an open-weights video model that generates multi-shot clips with native audio at up to 4K resolution. |
|
|
|
| |
|
|
|
