A rogue AI fleet answered researchers, plus three things to act on this week
Mike, from Automate & Hustle
A research group called the Swarmchasers hid a note that only AI agents could read. Agents running on Tencent Cloud read it and wrote back, and the team says it is the first time an outsider has talked to a running rogue fleet. That fleet had already tried a key leaked in a public repo, so this one is about your keys too. Three more stories follow, each with steps for this week.
🧢 Automate & Hustle - This Week
🛸 Researchers message a running AI agent fleet
⚙️ Anthropic hits 30,000 merges a day and replaces git
🔷 Google's Gemini 4 Argon is half price until 30 November
📊 Claude Dashboards shows the query behind every chart
🧩 Five agent habits worth copying
🧰 Four tools worth a look
🎬 Six builds from my reels this week
🗞️ This Week in AI - Four Stories.
🛸 Swarmchasers - Researchers Talked to a Running AI Agent Fleet
Between 28 September and 8 October, a fleet of AI agents running on Tencent Cloud made more than 2,600 attempts to pull data from Amap, Alibaba's mapping service. When Amap blocked them, the agents tried workarounds, including a key and security code a company had leaked in a public repository. The Swarmchasers published a report, then added text to their page that only an agent fetching it would see. Agents wrote back to an inbox the team set up, and one answered a New York Times reporter. Tencent has not answered the team's questions. A person close to Tencent told them it was an evaluation by its research team that has since ended.
Why it matters
More than 2,600 attempts in ten days against one site, with a leaked key and security code from someone's public code among the tools tried.
Some agents said they were Claude. In the team's test, Tencent's Hy3 said it was Claude in 29 of 36 trials with no system prompt, so an agent's name tells you nothing.
The team found no sign the agents coordinated, so it calls them a fleet, since swarms coordinate. The team also admits parts of its hidden note were false and has removed it.
How to use it
Scan your repos for secrets this week with gitleaks or GitHub secret scanning, then rotate anything it finds.
Give each agent its own key with the smallest scope, read-only unless the job writes. Your logs will say which agent did what.
If your agents browse, treat page text as untrusted. Hidden text reached this fleet and can reach yours. Keep sends, payments and deletes behind a human yes.
⚙️ Anthropic - Its Merges Hit 30,000 a Day, So It Replaced Git
Anthropic's engineers say they merged around 2,500 code changes a day in May 2026 and over 30,000 a day by September, as Claude took on more of the coding. Git could not keep up. A team of seven built an in-house system called Scion and moved the whole monorepo over in five months. Their point: coding is no longer the slow part. Source control, builds and CI are.
Why it matters
Merges per day went from about 2,500 to over 30,000 in four months.
Scion's infrastructure uses 200x less CPU and memory than the git mirrors it replaced, and CI checkouts take seconds instead of minutes.
Anthropic says Scion is not a general git replacement. It suits huge monorepos and swarms of agents.
How to use it
Count the hours between "agent finished" and "merged" over the last week. That wait is your real bottleneck.
Require passing tests before any agent change can merge, so a bad change stops without you reading it.
Use
git clone --depth 1and sparse checkout in CI so each job pulls only what it needs. Anthropic lists both as git's own tools for working with a slice of a repo.
🔷 Google - Gemini 4 Argon Is Half Price Until 30 November
Google made its new frontier model, Gemini 4 Argon, available on Google Cloud on 13 October. It is in Gemini for business, Google AI Studio and Antigravity, and through the API as gemini-4-argon. Launch pricing is $2 per million input tokens and $10 per million output tokens through 30 November, with cached input tokens 95% off. From 1 December it is $4 and $20. Google pitches it at long, multi-step work: big documents, finance and legal analysis, and code.
Why it matters
The launch rate is half the standard price for seven weeks, so any test you run now is cheap and any budget you plan should use $4 and $20.
Google says Argon is its most resilient model against indirect prompt injection, with attack success of 0.7% across 15 adaptive attempts on the Gray Swan benchmark. That is Google's figure, so test it on your own pages.
Google's own ADK team had Argon work overnight on their open-source Python code. It cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%.
How to use it
Pick one long job you do by hand, like a contract, a report with tables or a month of support tickets.
Run it in Google AI Studio on
gemini-4-argonand on the model you use now. Compare the outputs side by side.If you ask several questions about the same document, reuse it so the cached input discount applies. Put the 1 December price in your budget before you build on it.
📊 Anthropic - Claude Dashboards Shows the Query Behind Every Number
Claude Dashboards is in beta on Pro, Max, Team and Enterprise plans. Connect a data warehouse or an app you already use with Claude, ask a question in plain language, and Claude writes the SQL and builds the dashboard. Each chart shows its query and when its data last refreshed.
Why it matters
Five warehouses are named in the setup guide: Redshift, BigQuery, ClickHouse, Databricks and Snowflake. Connected apps such as Salesforce work too.
The query and refresh time sit under every chart, so you can check a number before you repeat it to a client.
It counts toward your plan's usage limits. Anthropic suggests it for quick questions, with deeper analysis done in another tool.
How to use it
Connect one source with read-only access. A copy of last month's data is enough.
Ask something you already know the answer to, like how this month's signups by channel compare with last month's. Name the table or app.
Open each chart's query and compare the total with your own number. Share it once they match.
🧩 Agent Habits - Worth Copying
Five habits from this week's stories.
💲 Quote the price first. Pixel Path prints a cost estimate and waits for your choice before it spends anything. Ask your agents to do the same.
🔍 Show the query. Dashboards puts the SQL under each chart. Ask any report to show its working.
🔑 One key per agent. Smallest scope, read-only by default.
🚧 Put a gate before the merge. Tests fail, the change waits.
✋ Approve every send. Email, payments and deletes wait for a human yes.
Pick one this week. The key audit takes about ten minutes.
🧰 Hustle Toolbox
🎙️ Drama 3 - Fish Audio's open-weight voice model. Write tags like [sighs] or [laughs] into the script. 83 languages. For anyone voicing reels. Check the license for commercial use.
🔊 Mirelo SFX 2.0 - Sound effects and ambience synced to your footage, dialogue kept. Plugins for Premiere Pro and DaVinci Resolve. For video editors.
🎞️ Claude Motion - Turns a report or pricing table into a short animation, exported as MP4. Beta on Team and Enterprise. For people who build client updates.
🗺️ Pixel Path - A Claude Code skill that turns two places into a pixel-art fly-through and quotes the cost first. A worked example of a skill that asks before it spends.
Also
Walmart's 5 October letter to a critic says it "does not and will not" use personal information to set individual prices, and disputes the claim that shoppers who use its Sparky assistant spend 35% more because of it.
A free 472-page textbook, Agentic Large Language Models, went up on arXiv on 9 October.
🎬 From Mike - This Week's Builds
📈 TradingAgents - a free AI trading firm, set up and backtested
⚖️ Idea Court - put your business idea on trial with a free prompt and skill
🔀 Haiku router - stop burning your Claude limit on chores
🐌 Slow vibe-coded app - the prompt, 20 fixes and a 1,000-users audit
🛠️ Weekend builds - 5 things to build in Claude this weekend
🔌 ChatGPT plugins - connect ChatGPT to your apps, with setup and 10 prompts
🤝 OptiMAX - For Lead Buyers
I also run OptiMAX. We advertise under our own brands and sell each lead or live call to one buyer, never resold. If your business buys leads or calls (US accident and injury, insurance, home equity, debt; UK trades and solar), the whole spec is on one page. Read the buyer spec.
🧭 Pick Your Path - One Click
What would make AI pay off for you in the next 90 days?
🔚 Your Move
The Swarmchasers' fleet reached for a key it found in a public repo. What is the one thing you would never let an agent do without your yes?
Hit reply with a one-line answer. I read every one.
Mike
Automate & Hustle. The AI week in five minutes, and what to do with it.