Hysen Labsgpt-pilot
The first real AI developer
GPT Pilot: an ambitious AI developer, with a serious security warning
GPT Pilot aims to be an AI that builds whole apps under a developer's supervision. The README now opens with something else first: a supply chain attack that went unnoticed for months, and what to do if you ran the code.
What the tool was supposed to be
The core technology behind the Pythagora VS Code extension. The pitch is a real AI developer companion, not an autocomplete: something that writes full features, debugs them, asks questions, and requests review. The stated research question is how much of a production ready app an LLM can build while a developer supervises. The README's own guess is that AI covers maybe 95 percent and the remaining 5 percent still needs a human until full AGI.
The security notice up front
A warning now sits at the very top. Malicious code was found and removed, a credential stealing worm hidden in the telemetry directory. A commit from August 2025, disguised as a routine revert, added a loader that downloaded the Bun JavaScript runtime and ran an obfuscated payload harvesting cloud and AWS keys, GitHub and npm tokens, and SSH keys. It was publicly reported in June 2026 and the malicious files were removed days later.
What it means for users
The payload only ran if GPT Pilot actually ran. Having a copy that was never executed is fine. Anyone who ran it from source between August 2025 and June 2026 should rotate every credential on that machine and check for signs of compromise, like an unexpected bun binary or specific files under the telemetry directory. The README is also clear that the repo is no longer actively maintained, and that this notice is a cleanup, not a return to development.
How it works, briefly
The README describes a pipeline of named agents, from a specification writer that asks clarifying questions, through an architect that checks which technologies are installed, to a developer and reviewer loop that passes tasks back and forth until they pass. The reasoning for going step by step instead of dumping a whole codebase at once is that bugs stay easier to fix when each piece is small and reviewable.
Editorial conclusion
The staged agent workflow is genuinely interesting, but the security incident is now the headline. The README is blunt about the timeline and about development having stopped.
Community notes