Mercor confirms breach via the LiteLLM supply-chain attack; Lapsus$ claims 4 TB of AI-training-contractor data

2026-03-31★★★policy-safetyMercorLiteLLMconfidence: medium

On March 31, 2026 Mercor, a $10B AI training-data and expert-contractor startup whose customers include OpenAI, Anthropic and Meta, confirmed a security incident. It said it was "one of thousands of companies" hit by a supply-chain attack on the open-source LiteLLM library. The extortion group Lapsus$ claimed about 4 TB of stolen data (source code, user database, interview videos and ID documents). Meta reportedly paused its Mercor contracts and contractors sued. The breach showed how one poisoned AI-infrastructure package can expose frontier labs' training-data pipelines.

Key facts

Mercor confirmed the incident on March 31, 2026 and attributed it to the compromise of LiteLLM, a widely used open-source library for calling LLM APIs (Fortune, TechCrunch)
The attack group TeamPCP planted credential-harvesting code in LiteLLM; secondary sources say malicious versions 1.82.7 and 1.82.8 were live on PyPI for roughly 40 minutes
Claimed by Lapsus$, not confirmed by Mercor: ~4 TB, including ~939 GB of platform source code, a 211 GB user database and ~3 TB of storage buckets with video interviews and identity documents (passport scans); samples published included Slack data, ticketing data and two videos of Mercor AI interviews with contractors
Mercor statement: it 'moved promptly' to contain and remediate; 'The privacy and security of our customers and contractors is foundational to everything we do'
Fallout (TechCrunch, Apr 9, citing Wired): Meta 'paused its contracts with Mercor indefinitely'; OpenAI was investigating its exposure but had not paused work; five contractors had filed lawsuits, one naming LiteLLM and compliance startup Delve as defendants
Context: Mercor raised a $350M Series C at a $10B valuation in Oct 2025 and was reportedly on pace for $1B+ annualized revenue before the leak
Re-reported on Oct 4, 2026 by TechJuice ('Mercor Breach: 4TB Data Leak Claims Shake AI Industry'); no new developments found in that piece

What happened

Late in March 2026 attackers tracked as TeamPCP compromised the publishing pipeline of LiteLLM, a popular open-source Python library that gives one interface to many LLM APIs. They shipped versions with credential-harvesting code. Secondary security write-ups say the malicious releases (1.82.7 and 1.82.8) were live for about 40 minutes and that the attackers reached LiteLLM's maintainer credentials through an earlier compromise of the Trivy scanner's GitHub Actions workflow. These details were not checked against a primary advisory. Mercor, which recruits domain experts to produce training data and RL environments for frontier labs, confirmed on March 31 that it had been hit. The Lapsus$ group then claimed it held about 4 TB of Mercor data and published samples.

Reports in early April said Meta paused its Mercor work indefinitely and that contractors whose IDs and interview videos may have been exposed had begun suing. Mercor did not confirm the size or contents of the stolen data, so the 4 TB breakdown remains the attackers' claim.

Why it matters

Data vendors like Mercor, Scale and Surge hold details of labs' secretive training projects (data specifications, labeling protocols, RL tasks) as well as sensitive personal data on large expert workforces. A single poisoned open-source dependency in the AI tooling stack was enough to put that at risk. The incident is pre-cutoff (March 2026) and is included as background for later AI supply-chain incidents.

Changelog

2026-10-04: created (sweep item: TechJuice re-report of Oct 4; facts from Fortune and TechCrunch, April 2026)

Related events

2026-09-22Snorkel AI raises $350M Series E at $3.5B as 'data-as-a-service' for frontier labs passes a $375M run rate ★★

Sources (5)

id: 2026-03-31-mercor-breach-litellm-supply-chain · updated 2026-10-04 · open in the interactive timeline