Muhammad Zulali, a YouTuber, says Anthropic’s Claude Code and OpenAI’s Codex read, then ran their own code inside, the engine computer of his 1994 Jeep Cherokee through a $3 Arduino Nano, and that the 32-year-old car then passed California’s smog test. His video, published on 10 November, had 1,417,084 views when we checked it on 15 November at 18:42 UTC, about 285,000 a day, by our arithmetic.
He says each smog test cost him about $100, and the parts list in his open-source xj-scanner repository comes to about $15, so one test costs about 6.7 times the whole kit, by our arithmetic. The video spread on X on 15 November, when @IntCyberDigest posted it to its followers.
Business Pill · AI Agents Explained
An evergreen Business Engineer explainer on AI agents, from answering to doing. It teaches the general idea only and does not discuss this video or the Jeep.
The key insight: As we read the video, the step that drew 1,417,084 views is the move from reading to acting: Zulali says Claude went from reading his 1994 Jeep’s engine computer to running its fuel pump, and Codex to running code inside it, through a $3 Arduino Nano and a kit of about $15.
Zulali says his Jeep was failing the smog test at 25 mph because of high NOx, and that a new oxygen sensor did not fix it: his 15 mph reading, which had been passing, went from around 80 to 900. He says each test at a smog station cost about $100, so he wanted the car’s own computer readings instead. The car predates the OBD2 diagnostic port of later models.
He says Claude Code worked out that the car talks over an old, slow protocol called SCI, and that a $3 Arduino Nano on a breadboard could act as the translator. After a resistor was pulled, the first reading came at 1:30 a.m., he says. A fuel-trim figure that showed 22% extra fuel turned out, on a later drive, to come from an online tool, ChryslerScanner, that he says read his car wrongly.
Costs Muhammad Zulali gives in his 10 November video and his xj-scanner repository: a $3 Arduino Nano, about $15 for the full parts list and about $100 for each smog test. The 6.7 times is our arithmetic.
From Reading to Running Code
The shift came when Claude suggested running the fuel pump itself, Zulali says, and he agreed. The pump stayed on for about 10 seconds, he says, until “I started to freak out and requested it to stop”. He describes it as the first time Claude acted on the car rather than only reading it.
Claude then copied the computer’s whole program to his laptop, he says, and found that his Jeep’s computer was a California version that trims fuel at the loads used in the smog test. A reflash to federal settings failed: his computer was made in December 1993, about four months before the programmable versions, he says. “So reflashing was physically impossible”, he says.
He says Codex found that the 1994 computer would run a small program sent over the same three wires, about 90 times a second, and then put the program on the EEPROM after the computer’s own code overwrote it in RAM. After many attempts, the car passed, he says, with Claude doing most of the diagnostic work and Codex the memory work. “I didn’t need a mechanic or like a Chrysler dealer tool for any of this”, he says.
YouTube views of Muhammad Zulali’s original video and three other channels retelling it, read on 15 November 2026 at 18:42 UTC. The 17 times is our arithmetic.
The Rules on Emission Controls
California’s Bureau of Automotive Repair says: “State and Federal anti-tampering laws specifically prohibit any modification to the vehicle’s original emission control system configuration as certified by the manufacturer”, citing California Vehicle Code section 27156. The Air Resources Board grants exemptions through an Executive Order: “If the part or modification is shown to not increase vehicle emissions, it is granted an exemption to emission control system anti-tampering laws”.
The video does not mention the Air Resources Board, an Executive Order or the exemption process, and it does not say what the code running on test day changed. Zulali’s video description says: “I’m not a mechanic. Some of what I say about engines may be wrong”. His repository calls the code a hobby project that “talks to your engine computer; you use it at your own risk”.
What Anthropic’s Usage Policy Says
Anthropic’s Usage Policy, effective 12 November, covers Claude Code and says users “are responsible for ensuring that agents they build or deploy, including actions those agents take through tools, browsers, or connected systems, comply with this Usage Policy”. It also bars using its products to “Violate any other applicable laws or regulations in your jurisdiction”.
For hardware, the policy says: “When connecting our models to hardware that can take physical actions, users should assess and understand” the possible outcomes if something goes wrong. Its extra requirements apply when outputs are “acted upon by hardware without human approval”, and its first category is equipment that can steer, accelerate, brake or navigate a vehicle. In the video, Zulali says he approved the fuel-pump run.
The repository tells owners of 1991–95 Jeeps and Chryslers to clone it and ask a coding agent to walk them through the hookup; it had 252 stars on GitHub on 15 November. For the market the two agents compete in, see our piece on Menlo Ventures’ AI coding count.
The Structural Read
As we read the rules alongside the video, the open question is not the code but the car: California’s Bureau of Automotive Repair says anti-tampering laws prohibit any modification to the certified emission control configuration, and the video does not say what the code on test day changed.
As we read Anthropic’s Usage Policy, responsibility sits with the user: it says users are responsible for actions agents take through connected systems, and its extra hardware requirements apply when outputs are acted on without human approval; Zulali says he approved the fuel-pump run.
Muhammad Zulali, in his YouTube video, 10 November 2026
“what else can you plug your agent into?”
Three Implications
FOR CODING AGENTS Zulali says Claude did most of the diagnostic work and Codex the memory work, while he did the wiring and the driving.
FOR CAR OWNERS IN CALIFORNIA BAR says modifications to the certified emission control configuration are prohibited unless exempted; CARB grants exemptions through an Executive Order.
The framework says of Anthropic: “It is helping define the rules, permissions, and boundaries around frontier intelligence.”
As we read it, the Jeep video shows where those boundaries meet the physical world: Anthropic’s Usage Policy puts responsibility for agents’ actions through connected systems on the user. The framework predates this video and does not discuss it.
What Is Not Established
The video does not state what the code that ran on test day did to the engine, and the words “fuel delivery” appear in the @IntCyberDigest post, not in the video. We did not see a smog test report, and the account rests on Zulali’s own video and repository. The transcript we read is YouTube’s automatic captions.
Whether the change falls under California’s anti-tampering rules is not established. Zulali, Anthropic, OpenAI and the state agencies were not contacted for this piece. This piece describes the video and does not explain how to repeat it.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
Muhammad Zulali says Claude Code and Codex, working through a $3 Arduino Nano, read and then ran code inside his 1994 Jeep’s engine computer, and the car passed California’s smog test. His video had 1,417,084 views in about five days, and one $100 smog test costs about 6.7 times his $15 kit, by our arithmetic.
A note on sourcing. Muhammad Zulali’s account is from his YouTube video of 10 November 2026, read through its description and YouTube’s automatic captions, and from his xj-scanner repository on GitHub; we did not see a smog test report. The rules quoted are from the pages of California’s Bureau of Automotive Repair and Air Resources Board and from Anthropic’s Usage Policy, effective 12 November 2026.
View counts were read on YouTube on 15 November 2026 at 18:42 UTC. The views a day, the cost ratio and the view ratio are our arithmetic. Zulali, Anthropic, OpenAI and the state agencies were not contacted for this piece, which does not explain how to repeat the work.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
