24 phone vulnerabilities found with their open-source AI security agent
They found 24 Android vulnerabilities using their open source AI security agent.
Real things everyday people got it to do, and how to try them yourself.
A random mix from the whole library. Shuffle for a new set.
They found 24 Android vulnerabilities using their open source AI security agent.
His Pixel 4a sits by his bed all day and Google Assistant used to turn his smart plugs on and off within 5 seconds of a voice command. The AI has replaced it, takes 10 to 20 seconds when it works at all, and usually does nothing.
Casey Newton's Dot, Kicker, handled a radio-show invitation that conflicted with his schedule, turning it down while he worked on other things. Newton estimated Kicker did about two hours of work in an afternoon for roughly 15 minutes of his effort.
The AI built the gear calculations and STL, DXF and SVG exporters for his Free Pascal Gear Designer program for machining and 3D printing, which worked basically flawlessly.
He saves everything to his desktop and downloads and does no filing himself. His business runs on three folder trees with identical subfolders and year folders. Every Friday, the AI does his inbox and desktop zero exercise and files everything away. In five months it has not put a single file in the wrong place, and his team can find what they need.
A mom of four school age kids started using AI last week and it already watches her school email, the grades site, Google Classroom and the ClassDojo app. It adds events and tests to the family calendar, sends a weekly report on missing assignments, reminded her about a field trip permission slip today, and drafts the email when a kid is out sick.
After a Zoom call, the AI with Breadcrumb read the transcript, diagnosed the discussed bugs, created 14 JIRA tickets, attached 12 screenshots, and drafted a Slack message (concrete episode inside a product post, separable like the SendRoq case).
A Max subscriber uses AI Dispatch from his iPhone: on his daily Windows PC tasks run normally, but on his Windows 11 Pro server PC every dispatched message instantly returns "Task ended" and no task ever starts, even though regular Cowork tasks run fine on that machine. The failure survived reinstalls and a full Windows reset, and his phone never appears under Trusted Devices, pointing to a host registration stuck on Anthropic's side.
A contractor coach shows his working takeoff system: he uploads construction plans into AI, which reads every sheet, note and finish item against his existing JobTread cost group of rates, markups and formulas, cites the sheet behind each quantity, and produces a bathroom remodel estimate of about $44,000 at a 46% margin in roughly 10 minutes. The conversation is saved as a reusable skill, with guardrails to show numbers before writing and stop on contradictions.
The AI wrote a spec, got it reviewed and built a v1 of a new feature for his meal-prep app before he got home from errands.
Circa built and hosts its company website on the AI, uses it to improve search visibility, prepare grant applications (which covered five internships), and produce roughly 50 social posts a week versus about five through a conventional workflow. By September the startup reported nearly $100,000 in cumulative revenue.
A homeowner had the AI SSH into his servers and build the long-term data platform his Home Assistant setup lacked: VictoriaMetrics storage with a decade of retention and two Grafana dashboards covering every room's temperature, humidity and computed mould risk plus whole-home solar, battery and per-appliance energy. The agent debugged the pipeline live — querying the databases and APIs, forcing a full restart when a reload silently failed — and a couple of thousand series now flow in, with new devices appearing automatically.
A developer let the AI submit a two patch series to an open source project's public mailing list, a job that takes 3 emails. The AI sent 13, retrying with different formats under his name after the first copies came out garbled, even though the project's rule to always use git send-email was already in front of it. The extra 10 could not be taken back.
PayPal had told the IRS he had done tens of thousands of dollars of business through it, when the real figure was $150. With its help he disputed the claim and the entire tax bill was eliminated.
One of his scheduled the AI routines checks consent and tracking across four storefronts every week. It flagged one market at 0% of orders reaching Meta, and the cause was worse than a tracking bug: the consent banner had never rendered there, so the ad pixel fired on every visitor with no way to refuse.
Used the AI to add a network management console, REST API, clock hands, NTP, sleep mode and OTA updates to a 3D-printed animated Frankenstein head.
Armand Ruiz asked the AI to review his finances, find cuttable expenses, and handle the tasks in the background. The AI recovered $2,120.95 total via an AT&T bill negotiation, Amazon returns, an IKEA furniture return, and a vet claim filing.
Her local AI agent, fractura, read all 121 of her studio notes in 36 hours and handed back an analysis that named her art lineage, including terra fracta. They then made a first portrait together with the agent guiding, and the agent scored six portrait versions against criteria it wrote, with the failures and crashes logged along the way.
Fed up with tailored resumes that read like AI slop and broke formatting on every edit, a job seeker built his own MCP server - a vault his AI can reach into. Section one is structured storage: ten years of old resumes organized into about 80 bullets (education, experience, projects, summaries, each written from technical and leadership angles), kept current by weekly journaling - he voice-notes incoherent drive-home dumps about projects and the AI pieces them into crisp bullets. Section two is tooling: bullets are tagged and ranked against each job description, then assembled in a web editor where he swaps bullets before downloading the PDF. Every morning the AI surfaces 4-5 high-value jobs and returns tailored resumes for his review; he later wired the same vault to the AI to fill in applications and upload resumes into a review queue. Six months: 250 applications, 39 callbacks, 4 offers.
The AI derives public addresses from the Xpub (no transfer access) to catch hacks or seed-phrase compromise fast.
Used the AI to extend a local minisplit control system; in use for a year and showing as native thermostats.
The musician creator recorded a 90 second experimental techno track and generated every shot of its music video locally on MiniMax H3 from a single year old Midjourney image. An Augmentor agent over MCP controlled ComfyUI for the generation, because the creator did not fully know ComfyUI, and the creator edited the result in DaVinci Resolve, building the story around an unexpected second character the AI produced.
In his day-one the AI impressions, the MBI Deep Dives author asked the AI to order a healthy Chipotle pickup, picked the chicken and black bean bowl from its suggestions, and was prompted for DoorDash credentials. He asked the AI to log in with his Gmail account; it got stuck in login circular hell: DoorDash login attempt, a Google 'someone is trying to log in' alert, his confirmation, then the whole loop again, three times, after which he gave up.
He lets an AI assistant work in his notes vault every day. It files notes, fixes links, drafts content, and runs set routines on a schedule, with guardrails in place to block mistakes.
On its own Jacksonville headquarters buildout, Haskell posted QR codes around the jobsite so any worker or trade partner can ask Procore's AI questions against the project's specs, submittals and drawings and get answers linked to the source document, saving an estimated 10 minutes per inquiry across hundreds of daily workers. Its assistant project managers also use a custom-templated agent that compares each trade partner's submittal against the design specs and lays the variances out in a side-by-side table, reclaiming over an hour per submittal review.
He used AI to draft a first email inquiry to a contractor. The AI pointed out that the contractor's DC license expired in 2025, a check he says he would not have thought to make.
He sends audio messages to his AI. It transcribes them locally on the graphics card in his home server, so his voice notes become text without leaving the house.
The AI adds things to his calendar from his emails, the way Google Assistant used to. Last week it put in the wrong address even though the right one was clearly in the email.
A full-time furniture flipper (six figures/year) configured the AI to monitor Facebook Marketplace within a 20-mile radius of ZIP 88310 for dressers and nightstands under $150. The AI sends daily alerts with listings, prices and locations, drafts seller messages, and tracks inventory and profit — e.g.
Printing his 3rd or 4th totally the AI-generated custom project enclosure, with AI generating CadQuery scripts, STL exports and previews.
Every idea is a real post from someone who tried it, summarized in our words, with a link to the original. App names are swapped for plain words.