Today's thread
Strung together, last week's news says one thing: AI is taking over more of the hands-on work. The Gemini app now edits PDFs and documents inside the chat, and Vercel Labs' agent-desktop lets AI find and press buttons in desktop apps. Google's newly released EGM has a second AI scrutinize the design before an AI writes any code.
The same trend shows up on the other side. VOLTA, a piece of malware analyzed by Cisco Talos, hands stolen account data to AI models and has them rank what is worth money. Google said in its earnings report that ads on its AI search answer screen earn less than classic search. Whether you delegate work or buy ads, the math is changing.
VOLTA malware uses Claude to appraise stolen data
On October 31, Talos, the threat intelligence team of security company Cisco, published its analysis of VOLTA, an information-stealing malware for Windows. An infostealer is a program that quietly takes passwords and login data from an infected PC. VOLTA collects saved browser passwords and cookies (small files a browser keeps so a site remembers you are logged in), payment cards, crypto wallet recovery phrases, Microsoft account login tokens, cloud access keys, Wi-Fi passwords and even screenshots. Using the stolen logins, it searches Gmail for receipts and pulls food delivery order histories.
What is new is that AI, not a person, appraises the haul. VOLTA sends requests to OpenRouter, a broker service that reaches many AI models through one gateway, and uses Claude Opus 4.6 by default. If a reply looks like a refusal, it moves on to Kimi K2 and then Gemini 2.5 Flash, trying three models in total. The output is a report that sorts the victim's assets into low, medium and high by how easily they convert to cash, and even rates the victim's security habits.
If one employee PC is infected, every work email, cloud and payment account logged in on it becomes part of that appraisal at once. Talos recommends keeping credentials in hardware-backed stores and, if a leak is suspected, revoking login tokens and changing passwords. For teams with IT staff, it also suggests blocking or alerting on Chrome launched invisibly with the --headless option.
Every endpoint Talos observed belonged to third-party services; it found no infrastructure run by the attackers themselves. The published analysis does not show how widely the malware has spread. There are also no public figures on how well AI companies' refusal safeguards block such requests.
Google Search revenue up 12% — AI Mode ads earn less
On October 28, Alphabet reported third-quarter results: Google Search revenue rose 12% to $63.35 billion from $56.57 billion a year earlier. Total revenue grew 25% to $127.6 billion, and Google Cloud revenue reached $30.71 billion, more than double the year before.
On the earnings call, Chief Business Officer Philipp Schindler said fast growth in AI Mode queries weighed on Search revenue because monetization in AI Mode was lower. AI Mode is a Google Search view that answers in conversation instead of a list of links. CEO Sundar Pichai said query growth was the highest in the last five years. In other words, queries are rising, but those that move into AI Mode carry fewer ads.
Google said it is scaling ads in AI Overviews and AI Mode to more users and queries and expanding AI Mode ads to international markets. Schindler said click-through rates for sponsored and organic product listings in AI Mode are already similar. For businesses running Google Search ads, this signals that ad placement is moving from link lists into AI answers, and that whether product information is captured accurately in those answers matters more.
Schindler did say currency movements were the single largest driver of Search deceleration versus Q2. Google disclosed no AI Mode ad pricing or revenue, so the size of the gap is unknown. This report also does not confirm when AI Mode ads will open in Korea.
Gemini app now edits PDFs and docs inside the chat
On October 29, Google let the Gemini app create, preview and edit documents, PDFs, presentations, spreadsheets and code files inside the chat. You can expand a file to full screen and type changes directly, highlight a passage to ask about it, or ask Gemini to refine it. Google said the feature is rolling out to all users in every language and region where the Gemini app is available.
In April, Gemini gained the ability to generate downloadable files such as PDF, Word and Excel from a single prompt. Now the editing step after creation has moved into the chat as well. Previously you had to take the file into Word or a PDF editor to fix it.
For documents where AI drafts and a person touches up a few spots, such as meeting notes, notices or quotes, there is less switching between programs. PDFs can be opened and edited on both web and mobile, but Google Docs, Sheets, Slides, Markdown, text and code files work only on the web. On a phone, PDFs are the practical choice.
The announcement lists no differences by plan or usage limits. It is also unconfirmed whether admins can turn the feature on or off for company accounts. How well original formatting survives is something to judge by opening a file yourself.
Google EGM — a memoryless AI reviews designs before coding
Google released Elephant-Goldfish Model (EGM), a plugin for AI coding tools, as open source under the Apache 2.0 license. A plugin is an add-on package that extends an existing tool, and open source means software whose code anyone can read and reuse. An elephant AI that remembers the full conversation writes the design doc, and goldfish AIs that know nothing of the earlier conversation read only that doc and look for gaps. A person weighs the issues and approves the design before handing off implementation.
Until now, people often asked the same AI, in the same chat where the plan was made, to review that plan. The AI then takes what it already knows for granted and easily misses what the document leaves out. EGM is built to stop the review entirely if the environment cannot spin up separate AI workers that share no memory with the main task.
Supported environments are Google Antigravity, Claude Code and Codex, with setup guides for Claude Cowork and ChatGPT Work. In Claude Code, install it with claude plugin marketplace add google/elephant-goldfish-model followed by claude plugin install egm@egm; a new session then gets commands starting with /egm:. As of November 3, the GitHub repository has about 10,700 stars.
The repository recommends it for features that need design decisions, not for small, clear bug fixes. It also notes that review does not replace tests or human judgment. It carries the caveat that it is not an officially supported Google product.
Vercel agent-desktop — drives apps by UI element, not coordinates
agent-desktop, built by Vercel Labs, the experimental arm of web hosting platform company Vercel, is a command-line tool that lets AI read and operate desktop apps. Instead of taking screenshots and clicking coordinates, it finds elements in the operating system's accessibility tree. The accessibility tree is the structured list of on-screen items that screen readers use to read out button and field names. It runs on macOS, Windows and Linux under the Apache 2.0 license.
Clicking based on screenshots easily hits the wrong spot when window size or resolution changes. Each time agent-desktop takes a snapshot, it returns a reference tag for every element that can be acted on; the AI acts on that tag and then takes a new snapshot to check the result. By default it neither moves the mouse pointer nor steals window focus.
It is worth trying if you want AI to handle repetitive entry in in-house programs that have no web integration. On a PC with Node.js, one line, npm install -g agent-desktop, installs it; on a Mac, the terminal needs Accessibility permission. As of November 3, the GitHub repository has about 9,900 stars.
On Linux, some features differ depending on the desktop environment, according to the documentation. Because the AI presses real buttons in real apps, it is safer to start with tasks whose results a person can check right away.
Claude Cyber Verification Program opens only to security defense
According to Anthropic's help center, the Cyber Verification Program (CVP) gives security professionals and defense-focused organizations advanced cyber capabilities and fewer blocking safeguards that halt work midway. Covered models are Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models. There are three access tiers, Defense, Red Team and Specialized, and individuals on paid plans can apply only for Defense access.
In standard Claude, tasks such as malware analysis or vulnerability validation can be stopped by safety classifiers. CVP is the route for people doing that security work to use Claude after verifying identity and purpose, and applicants hear back by email with a decision or a request for more information within seven business days. On the Pro plan, Mythos 5.1 runs only on separate usage credits rather than included usage, and on Max plans Mythos 5.1 and Fable share a single allowance of up to 50% of the weekly limit.
It is not a workaround for non-security automation that gets blocked. The documentation says organizations whose main business falls outside defense may not qualify. If a company's security staff keep getting blocked while checking malware with AI, the structure is for the organization to apply once and for admins to assign seats.
Defense access must adopt phishing-resistant multi-factor authentication and stop using API keys by December 15, 2026. Multi-factor authentication adds a second check, such as a security key, on top of a password, and an API key is the secret key a program uses to call Claude. Until then, some form of multi-factor authentication is already required and API keys expire every seven days, so approved teams should change how they sign in before that date.
Try this today
Open the list of work passwords saved in your browser and move payment, cloud and email accounts first to a password manager or passkeys (logging in with your phone's fingerprint or face recognition instead of a password). Ending login sessions on devices you no longer use, from your account settings, also shrinks what malware like VOLTA can take.
If you run Google Search ads, write down this month's impressions, clicks and conversions next to last quarter's. It becomes your baseline for seeing what changes as ads inside AI answers expand.
On Gemini's web app, create one PDF notice you plan to send this week and fix just two spots through the chat. Compare it with the original for broken formatting before deciding whether to use it for real work.
If you plan to build your next feature with Claude Code or Codex, install EGM and run one design doc review. If the goldfish AI raises questions you cannot answer, settle them before implementation.