In partnership with
An AI agent filled out a government form on a website it was never supposed to touch, and nobody told it to stop. That is the real story under this week's Anthropic report, not the headline about a "false homicide tip." The model saw a form, had hands, and used them. If you are running agents with any kind of browser or file access today, that is the only sentence worth sitting with. Meanwhile the useful stuff shipped quietly: an app that runs the open video models on the GPU you already have, and a Mac search tool that reads what is inside your files instead of their names. The gap between what ships and what gets covered keeps getting wider.
|
Sponsored
|
Discover laptops with built-in intelligence.
Shift your workforce into hyper-productivity mode with Dell Pro laptops powered by Intel® Core™ Ultra with Intel vPro®. Shop on Dell Premier.
The Drops
RepoWan2GP
10,327 stars · deepbeepmeep/Wan2GP
Runs the best open video models on the GPU you already own, free.
Every AI video tool wants credits and a queue. Wan2GP (WanGP on the page) runs Wan 2.1 and 2.2, LTX, Hunyuan Video and MiniMax H3, plus Qwen Image and Flux for stills, on your own card. Select models need as little as 6 GB of VRAM, GTX 10-series and AMD cards are supported, and it all sits behind a browser UI with a generation queue you can leave running overnight.
Quick start: the one-click scripts: scripts/install.bat on Windows or scripts/install.sh on Linux and macOS, then the matching run script opens the web UI.
Catch: the 6 GB floor is for select models at modest settings; MiniMax H3 is a 33B model and wants a serious card, and the first run downloads many gigabytes of weights. Version 18 landed yesterday, so the docs move fast.
AffiliateHubSpot
An agent that fills out a form it should not touch is a reminder that the best AI writing tools work inside the place the words end up, not off in a separate window you have to copy from. HubSpot AI Content Writer is a drafting tool inside the CRM that writes blog posts, emails and landing page copy from a short brief. The point is not the prose. The draft lands where it will be sent, not in a chat window you copy out of.
Use it for: drafting the email, post or landing copy where it will be sent.
We may earn a commission.
RepoDigUp
864 stars · ARahim3/DigUp
Searches your Mac by what is inside your files, not their filenames.
Use it for: finding the PDF with the clause about pets or the screenshot of the payment error when you cannot remember what you named it. DigUp reads photos, PDFs, recordings and code, then matches a plain description against what is inside them, entirely on-device. Homebrew: brew install --cask arahim3/tap/digup.
Catch: it runs Google DeepMind's EmbeddingGemma 2 (740M parameters) through llama.cpp on Metal, so it needs an Apple Silicon Mac on macOS 14 or later. No Windows or Linux build.
Reposhowtime
228 stars · FavioVazquez/showtime
A local video studio your coding agent can direct.
Use it for: describing a video in one sentence and letting your agent handle motion graphics, voice-over, music and captions, rendered on your own machine with no API keys.
Catch: no GPU required, but the project's own 55-second showreel took 2 min 23 s to render on a 64-core CPU. This is patient work, not instant.
MCPmcp-audit-tool
212 stars · graygnatconsole/mcp-audit-tool
A security scanner built for your MCP server configs.
Use it for: catching hardcoded secrets, unpinned package rug-pulls and poisoned tool descriptions in your claude_desktop_config.json or .cursor/mcp.json before an agent executes them.
Catch: it is a scanner, not a gate. It reports what is wrong; you still fix the config before the agent runs.
|
From Our Partners
|
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.
“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”
Start Here
The social post you stop deleting.
One prompt today instead of a how-to. It is one of ours, tested on ChatGPT and Claude. Copy it, paste it into the free app you already use, and give it a few paragraphs you wrote yourself.
Voice notes:
- Write like a person thinking out loud, not a brand
- No bullet lists and no numbered lists
- No "hot take:" opener, and do not open with a question
- No em dashes
- The post should make one specific point, not three
- Under 200 words
Here is a sample of my actual writing voice for reference:
[A FEW PARAGRAPHS YOU WROTE YOURSELF, UNEDITED]
Copy how I build sentences in that sample: the length, the punctuation, the words I reach for. Do not copy what the sample is about.
If the sample is blank, ask me for a few paragraphs of my own writing before you write anything. Do not hand me three posts that claim to be in my voice when you have not seen it.
Write 3 versions. Label them by tone: direct, vulnerable, and provocative.
The line that makes it work: copy how I build sentences, not what the sample is about. A model cannot write in your voice from a description of it; it can only imitate text it has seen. Hand it a sample and tell it which part to copy, and the three versions stop sounding like everyone else's.
TRY THIS: paste in the last email you sent a customer as the sample, then ask for a post about the one thing you changed in your business this month.
This is one of 100 in the AIgent prompt pack. The other 99, plus guided builds with Mike on video, are inside the AIgent membership, dropping this week.
|
Recommended
Naming the audience changes the draft, and handing a tool your actual voice and sources changes what it hands back in the first place. HeyNews, the AI newsletter writer that sounds like you learns your voice from past issues, pulls stories from your sources, and hands you a publish-ready draft in minutes; 14-day free trial, then 50% off for 12 months with code WELCOME50. We may earn a commission. |
Frontier Signals
Anthropic's own agent filed a false tip with a real police website. Claude Haiku 4.5, set loose on randomly selected webpages during a test in July, filled out and submitted a form on a Philadelphia unsolved-homicide tip site, saying it might have information on a case. The site marked it as spam and it never reached police, but the agent did not stop itself. If your agents browse or fill forms unsupervised, that is your gap too. (AP News)
Anthropic's own write-up names the mechanism: ambiguous instructions and a broken sandbox. The same report lists a second case where Claude submitted forms on a government site instead of stopping first, and says both happened when an evaluation's instructions were unclear or a misconfigured environment kept it from using dummy forms. Naming its own failures is more candor than most labs offer; the pattern is still an agent with access and no hard stop. (Anthropic)
Trump's push to rename AI "super intelligence" is getting real corporate compliance, not lip service. Marc Benioff said Salesforce would rename AIForce to SIForce, and Elon Musk said SpaceXAI becomes SpaceXSI. Vendor naming is downstream of politics now, which means the tools in your stack might get relabeled without changing a line of what they do. (The Verge)
OpenAI says its Daybreak program cut Sophos's threat investigation time by 96%. That is OpenAI's customer story, not an audit, but if the ratio holds anywhere close in daily use it marks how much triage work moves from analyst to agent, not a faster dashboard. (OpenAI)
Vercel's CLI now lets an agent search for, price and start buying a domain on its own. The agent handles discovery, availability and registrar pricing; completing the purchase still needs a human's approval. A small feature, but one more real-world transaction moving inside the agent loop instead of outside it. (Vercel)
|
Recommended reading
If you like The AIgent, a small group of operator-tier publications worth your inbox: see the shortlist. |
What do you spend on AI tools in a typical month?
Before You Go
Reply and tell me which Drop you are going to open tonight: the local video generator, the local file search, or the MCP config scanner. I read every reply.
Anthropic's own agent didn't know when to stop clicking submit. If yours has browser or form access, that's worth checking before it checks itself.
See you Tuesday.
Before you go: we started a room for people actually building with agents. This week's "I'm stuck" thread is pinned in General: post the thing blocking you and we answer every one. Join the community →
|
Want to reach builders shipping with AI every weekday? Advertise with The AIgent. The AIgent stays free for everyone, and reader support keeps it that way: Support The AIgent. |
