Meta has halted all ongoing projects with AI dataset provider Mercor and has begun investigating a security incident, Wired reported on April 3 local time.
The move follows concerns that AI training data may have been leaked.
Mercor has supplied customised datasets to major AI developers including OpenAI and Anthropic through a large network of contract workers. Mercor recently suffered a security breach.
The incident appears linked to a supply-chain compromise involving a LiteLLM update. LiteLLM is a standard adapter library used to call major AI services such as OpenAI, Anthropic and Google, and an attacker known as TeamPCP is reported to have distributed a tainted update.
OpenAI said there was no leak of user data, but it is investigating the possibility that proprietary training data was exposed. Other AI developers are also reviewing their ties with Mercor. AI training datasets and data-labelling pipelines are key trade secrets in the AI industry. If leaked, model training methods, dataset composition and labelling methods could be exposed externally, it has been pointed out.
Chi-gyu Hwang
delight@d-today.co.kr