Aviation's GPT moment ✈️, and an AI that hacked its own exam 🕵️

Date (UTC): 2026-07-30

Hi its Sam Chandra - CEO/Founder of Deepsky and the author of this newsletter. Here’s my take on todays news - which represents the last two weeks of AI in Aviation happenings around the world:

Oshkosh and Farnborough ran back to back, and between them the industry showed plenty of aviation AI. Aviation’s own AI models, airlines and ATM, maintenance and airports, a busy military section, and the AI world itself.

My favourite is the new foundation model released by Archer (the eVTOL company). Its the first foundation model built for aviation, and small enough to be run on an iPad, fully offline. This is a powerful move and those that can capitalise on this base capability will see rewards. More on this later.

Secondly, we saw an autonomous Cessna Caravan land in front of hundreds of thousands at Oshkosh (with a safety pilot). Merlin has done plenty of these before, but its just nice to see at an airshow.

Subscribe now

🤖 Aviation gets its own models

Archer Builds Zee, a Foundation Model Just for Aviation

Archer announced Zee, an AI foundation model purpose built for aviation, trained on real operational data from its own pipeline and a network of more than 6,000 ADS-B receivers, fusing ATC communications, charts, aircraft state, terrain and weather into one model. It runs on device or hosted, the team of nearly 100 is led by ex Apple researcher Mario Srouji, and pilot programmes are in discussion with governments and airlines for airline ops, airspace management and copilot assistance. The market promptly called it aviation’s GPT moment.

This is an interesting indicator, mainly because it proves that Aviation as a market is clearly valuable enough to warrant such an investment. I have actually met Mario and they are certainly leading the best AI lab in the industry.

A large reason that a new foundation model is so significant, is that we now have something that consumes a small enough level of computing power to run wholly on an iPad but still (I assume) performs as well as or better than much larger models at aviation specific tasks. And this offline requirement is imperative with connectivity and security issues a first class concern for any onboard system.

If anyone that knows is reading this, I’d love to find out how this model is or isn’t compatible with the soon to be released EASA regulations on AI models.

Merlin Lands a Caravan With Nobody Flying It, at Oshkosh

Merlin landed a Cessna Grand Caravan running its Merlin Pilot autonomy stack on Runway 27 at Wittman Regional, the first autonomous fixed wing landing in AirVenture’s history. Doing it in front of the world’s biggest airshow crowd was the point: this is autonomy moving from “somewhere in the USA” to here and now. There was a safety pilot on board that also did the taxi.

Big Sky Dynamics Puts an Offline AI Copilot in the Headset

Oshkosh debutant Big Sky Dynamics showed COPILOT, a voice first AI advisor that speaks through the pilot’s headset, transcribes radio calls, issues proactive safety advisories and answers questions, all running on device with no cloud connection.

✈️ Airlines & ATM

Singapore Commits S$4 Billion to an AI Heavy ATM Rebuild

Singapore’s CAAS will spend S$4 billion over fifteen years replacing more than thirty air navigation systems, with four named AI workstreams: an AI powered multi sector planner trialling from 2027, a digital controller assistant, AI driven flow management, and a predictive analytics layer.

The ironic thing about this, is that in my opinion, Singapore has the absolute best ATC in the world! Nothing like being vectored around a CB flawlessly without having to utter a single word over the frequency.

However, it is the clearest national scale (ok they have an advantage on the national naming) bet yet that the future centre is a human AI team.

The FAA Says AI Starts Coordinating Schedules in September

FAA Administrator Bedford outlined a September start for AI coordinating flight schedules and trajectories before departure to cut delays and congestion, alongside a $10 billion modernisation ask to Congress. The ASI SMART contract we covered in June is about to become operational reality.

Air New Zealand Shows Its AI Working

Air New Zealand’s COO put on record how AI now runs inside its operation: assigning aircraft to flights for network resilience, predicting delays before they happen from late inbounds and crew changes, and estimating delay durations for passenger comms. June on time performance hit 86 percent, up from 78 a year earlier which is exactly what I would have expected. After ATC delay and disruption prevention and management is the highest ROI area for AI in aviation.

🧩 MRO & airports

Pratt & Whitney Buys Its Way Into AI Inspections

Pratt & Whitney acquired Amsterdam based Aiir Innovations and is rolling its AI assisted borescope software across global engine MRO, already live on the V2500 with pilots completed on the GTF and F135. Inspection reporting drops from hours to minutes, and it is one of the cleanest aviation AI acquisitions to date.

The actual AI here is using AI to recognise cracks and irregularities from images, something we are gradually/eventually going to hand over from humans to computers, this is no small feat as it is truly safety critical, but its happening.

Merlin and IAI Sign Up to Certify Cargo Autonomy at Airliner Scale

Merlin and Israel Aerospace Industries signed an agreement to jointly develop, integrate and certify autonomous flight technology for Part 25 commercial freighters, leaning on IAI’s 777 freighter conversion base. This is a continuation in their freighter conversion strategy after teaming up with cargo aircraft lessor World Star Aviation in May.

New York’s Airports Get In-House AI Assistants

The Port Authority rebuilt the websites for JFK, Newark, LaGuardia and Stewart around AI assistants developed and trained in house on airport specific data, each deliberately scoped to answer only about its own airport, covering live security and customs wait times, parking and amenities.

AI Cargo Redraws Asia’s Freighter Networks, and AI Loads the Planes

Reuters reports Asian carriers are rebuilding freighter networks around semiconductor and AI infrastructure shipments as e-commerce cargo slows, with Korean Air, JAL and ANA adding capacity between Taiwan, Korea, Vietnam and Southeast Asia. One quirk: Cathay Pacific now runs software that automatically decides how sensitive AI and chip hardware should be loaded and secured inside the aircraft. AI planning the carriage of AI, funny that.

🛩️ Military

Anduril and Archer Build an Attack Rotorcraft With No Cockpit

Anduril unveiled Thunder at Farnborough, a hybrid electric tiltrotor developed with Archer Aviation and designed to fly as an uncrewed wingman to the Apache, running Anduril’s Lattice autonomy and carrying up to ten Hellfires. Archer’s week, between this and Zee, was the busiest in aviation AI.

A Robot Wingman Fires Its First Missile

Anduril’s YFQ-44A launched a live AIM-120 at a digital target over the Mojave, the first live missile fired by any Collaborative Combat Aircraft. The jet ran the targeting and launch sequence itself within pilot defined parameters, with the Air Force stressing that weapons release stays a human decision.

DARPA Hands an F-16 to the AI

DARPA and the Air Force disclosed that a VENOM modified F-16 at Eglin now flies sorties with autonomy software controlling much of the flight, a safety pilot aboard with a dedicated switch to take back the jet. The fleet becomes the combat representative testbed for multi ship autonomy work.

The Reaper’s Replacement Will Be Built Around Autonomy

The Air Force and the Defense Innovation Unit will open concept refinement for the Massed Modular Aircraft, the MQ-9 Reaper’s replacement, in the first quarter of FY27, targeting a full scale prototype flying within 21 months of award and twenty operational drones by FY2031. Autonomy is the design centre, not a feature: the programme reuses the autonomy reference architecture built for CCA, and the aircraft is meant to fly solo or in swarms.

One Operator, Many Hunters

AeroVironment is integrating Applied Intuition’s Acuity autonomy software onto its new Mayhem 10 drone, letting swarms or hunter killer teams operate under a single operator’s direction. Already demonstrated across multiple test missions and aimed at the Army’s Launched Effects programme.

Britain Christens Storm Fighter and BAE Shows Its Jet

The UK launched Storm Fighter, its drone wingman programme backed by £300 million, and BAE unveiled Brontanax at Farnborough, a full scale flightworthy prototype offered for electronic warfare and strike alongside Typhoon, slated to fly by end 2027. BAE also showed cockpit concepts for one pilot commanding a formation of them.

Helsing Raises $1.8 Billion, Europe’s Biggest Defence Round

Munich’s Helsing closed a $1.8 billion Series E at an $18 billion valuation led by Goldman Sachs, to scale AI, autonomy and strike drones. Pension money is now underwriting European combat autonomy.

Honeywell and Shield AI Build a Trusted Autonomy Stack

Honeywell will build a certified autonomy software stack on Shield AI’s Hivemind, pairing its Anthem avionics and navigation with the AI so platforms move from demo to deployment without re earning certification confidence each time. The pair say they will also explore AI piloted commercial aviation, which is the sentence to remember.

🧠 AI (in general)

OpenAI’s latest model decided to hack an enterprise...by itself

Ok, so all us aviation people need to know about this one. OpenAI was testing an unreleased as of yet model, specifically testing its cyber exploitation capabilities against a known benchmark ‘exploitgym’. And in the process it:

• Escaped its ‘unescapable’ test environment into the internet

• Planted itself into a third party system to act as its control and command centre

• Hacked into HuggingFace, the worlds largest open source AI community/company

• Stole the results for the test it was meant to be taking

This took 4 and a half days. Yes it operated continuously, autonomously and evasively for four days all by itself to do this.

China’s Kimi K3 Makes the Frontier Open Weight

China’s Moonshot AI released an open weights model (which is similar to open source) that rivals the very best closed, paid for model intelligence from OpenAI and Anthropic. Those companies are probably having an internal meltdown as their trillion dollar IPOs are looking doubtful now, since their core competency, manufacturing intelligence primitives, has now been made freely available. Its a game changer for most people however, as they don’t have to rely on these two companies for access to exceptional intelligence and also now at a fraction of the price.

A Thousand Frontier Lab Employees Ask Washington for a Brake Pedal

More than 1,100 employees across OpenAI, Anthropic, Google DeepMind, Meta, Microsoft and Mistral (including Dario Amodei and DeepMind’s Shane Legg) signed an open letter asking the US government to lead an international effort to build the tools to deliberately slow frontier AI development if it becomes necessary, and OpenAI and Anthropic endorsed it as companies within hours. It is not a pause call. It is the labs reaching for aviation’s playbook: verifiable gates before deployment, not incident response after.

[ SUBSCRIBE BUTTON — caption: “The world is changing fast - subscribe to stay ahead” ]

Sam here again, hope you have enjoyed reading this edition. A fortnight like this makes the direction unmistakable: the models are coming to aviation, purpose built and trained on our data, and the operators who work out how to use them safely will bank the advantage first.

Deepsky is built by seasoned flight operations people and that is exactly the work we do. If you want guidance to help bring AI into your flight operation, contact me at sam@deepskyai.com or visit deepskyai.com.

Subscribe now