AI Training Data Breach Hits OpenAI and Meta; Foxconn Q1 Surges 30%

A supply-chain hack exposed AI training secrets at OpenAI and Meta as Foxconn's Q1 revenue jumped 30% on AI hardware demand.

🔐 LiteLLM Supply-Chain Attack Exposes AI Training Secrets at OpenAI, Meta, and Anthropic

Decoded: A cyberattacker group known as TeamPCP compromised two versions of LiteLLM — an open-source AI API proxy used across the industry — planting malicious code that exposed downstream companies using the library. Mercor, one of the primary contractors that generates proprietary training data for OpenAI, Anthropic, and Meta, confirmed it was among the victims in an internal email to staff on March 31. Meta has indefinitely paused all work with Mercor while investigating. OpenAI is investigating the incident but has not suspended its Mercor projects. Mercor's business is creating bespoke human-generated datasets used to train models like ChatGPT and Claude — datasets kept highly secret because they reveal competitive details about how each lab trains its AI. Whether the exposed data was accessed by a state actor or competitor has not been publicly confirmed. (Wired, April 3, 2026)

Why it matters: AI training data is one of the few genuinely proprietary moats left in the model race — the difference between open weights and competitive advantage increasingly comes down to the quality, structure, and sourcing of training datasets. A breach at a vendor serving OpenAI, Anthropic, and Meta simultaneously is structurally significant: if training data details were exfiltrated, competing labs — including those in China — could learn specifically how leading U.S. models are curated, annotated, and filtered. LiteLLM is widely embedded across the AI industry as a proxy layer for routing API calls; a supply-chain compromise via an open-source library is harder to detect than a direct intrusion and exposes thousands of downstream users. For Meta (META), the indefinite pause signals the breach crossed a materiality threshold internally. The incident highlights a systemic risk the AI industry has not yet addressed: critical training data infrastructure is concentrated in a small number of contractors operating with minimal security requirements.

📊 Foxconn Q1 Revenue Jumps 29.7% on AI Server Demand — Flags Tariff Uncertainty

Decoded: Taiwan's Foxconn — the world's largest contract electronics manufacturer and the primary assembler of Nvidia AI servers — reported a 29.7% year-on-year rise in first-quarter 2026 revenue, driven by strong demand for AI products. The company flagged "volatile" global politics and trade policy uncertainty as near-term risks — a reference to U.S. tariff escalation that accelerated through Q1. Foxconn assembles Nvidia's H100 and B200 server systems and builds hardware for Apple, Google, and major hyperscalers. Q1 2026 results were released April 5. (Reuters, April 5, 2026)

Why it matters: Foxconn's 29.7% Q1 growth is a direct readthrough for AI hardware demand — the company sits at the intersection of Nvidia GPU supply and hyperscaler server assembly. A near-30% revenue increase validates that AI buildout spending has not slowed entering Q2 2026. The tariff flagging is material: Foxconn assembles in Taiwan, Vietnam, India, and Mexico — all in scope for potential U.S. tariff pressure on electronics imports. For Nvidia (NVDA), tariff-driven cost escalation on AI server hardware is an emerging downstream risk on the infrastructure buildout thesis. The Q1 result also complicates the bear case: actual shipped hardware volume is growing regardless of what equity valuations are doing in the tariff-driven selloff.

Stay decoded. See you tomorrow.

— The Get AI Decoded Team