Tesla ex-VP shares AI coding horror that could have wiped sensitive data; Musk agrees

1 month ago 24
ARTICLE AD BOX

Tesla's erstwhile  VP Andrej Karpathy shares AI coding 'horror', 'Python proviso   concatenation  attack' that could person  wiped millions of SSL backstage  keys, database passwords, more; and Elon Musk 'agrees', says ...

Andrej Karpathy, the erstwhile Tesla AI manager and OpenAI cofounder, is calling a caller Python bundle onslaught "software horror"—and the details are genuinely alarming. A compromised mentation of LiteLLM, 1 of the astir downloaded AI libraries connected PyPI with 97 cardinal monthly downloads, concisely turned a regular pip instal into a credential theft cognition susceptible of exfiltrating SSH keys, AWS and Google Cloud credentials, Kubernetes configs, crypto wallets, SSL backstage keys, CI/CD secrets, and afloat ammunition histories.The malicious versions—1.82.7 and 1.82.8—were uploaded straight to PyPI connected March 24, bypassing LiteLLM's authoritative GitHub merchandise pipeline. The onslaught was traced to TeamPCP, a menace histrion connected a multi-week run done developer and information tooling. They had already compromised Aqua Security's Trivy scanner days earlier, which gave them entree to LiteLLM maintainer BerriAI's PyPI people token.

How a bug successful the malware really saved thousands of developers

The poisoned bundle was unrecorded for astir 2 hours earlier PyPI quarantined it—and the lone crushed it got caught that accelerated was a mistake successful the attacker's ain code.

Developer Callum McMahon was installing a Cursor MCP plugin that pulled LiteLLM arsenic a transitive dependency. Version 1.82.8 caused his instrumentality to tally retired of RAM and crash. That clang acceptable disconnected the alarm. "If the attacker didn't vibe codification this attack," Karpathy wrote connected X, "it could person been undetected for galore days oregon weeks.

"

Karpathy says the incidental is simply a crushed to rethink however developers usage dependencies

Karpathy utilized the incidental to revisit a long-standing concern: that the bundle industry's reliance connected dependency trees creates enormous, mostly invisible onslaught surfaces. Every bundle successful a project's concatenation is simply a imaginable introduction point. His suggestion—increasingly his default—is to usage LLMs to extract oregon replicate elemental functionality alternatively of importing full libraries. Maintainers astatine BerriAI person since engaged Mandiant for probe and advised contiguous credential rotation crossed the board. Docker images, which pin dependencies, were confirmed unaffected.

Read Entire Article
LEFT SIDEBAR AD

Hidden in mobile, Best for skyscrapers.