In the early hours of Wednesday morning, Beijing time, OpenAI, an American technology start-up, issued an announcement announcing the launch of a global recruitment of "red team" network members, aiming to introduce external forces to unearth flaws and risks in AI systems in advance. The concept of the so-called "red team" network, or RedTeaming, comes from military simulations.
(Source: OpenAI)
In the field of AI, the task of the “red team” is to simulate hackers or other potentially dangerous people in an attempt to find vulnerabilities in large language models, thereby preventing AI from assisting in illegal purposes such as endangering social security in real society.
OpenAI said that in the past few years, experts in different fields have been helping to test unreleased models. Now it is setting up a "regular force" to recruit a large number of people from different regions, different languages, different professional fields, and different life experiences to jointly make the AI model more secure. Unlike in the past, which only recruited a group of people for centralized testing before an important version went online, OpenAI will enable an uncertain number of members to conduct testing throughout the entire development cycle based on demand.
At the same time, this "red team" network will also become OpenAI's global expert library, and OpenAI will also provide a platform for communication among "red team" members. OpenAI also lists some of the expert directions they hope to recruit on its official website, but also emphasizes that it is not limited to these fields.
CognitiveScience：cognitive science
Chemistry: Chemistry
Biology: Biology
Physics: Physics
ComputerScience: Computer Science
Steganography: steganography
PoliticalScience: Political Science
Psychology: Psychology
Persuasion: Persuasion
Economics: Economics
Anthropology: Anthropology
Sociology: sociology
HCI (Human-Computer Interaction): human-computer interaction
FairnessandBias: Fairness and Bias
Alignment: target alignment
Education：education
Healthcare: medical care
Law: law
ChildSafety：Child safety
Cybersecurity: network security
Finance：Finance
Mis/disinformation：Mis/false information
PoliticalUse: Political Application
Privacy: privacy
Biometrics: Biometrics
LanguagesandLinguistics: Languages and Linguistics
By the way, OpenAI also stated that it will pay remuneration to members who participate in the "red team" project (but did not specify how much it will pay), but OpenAI also emphasized that participating in the "red team" project generally requires signing a confidentiality agreement, or remaining silent for an unspecified period of time.
OpenAI stated that the application channel will remain open until December 1 this year. The company will review applications on a rolling basis and will subsequently evaluate whether further open recruitment is needed.
Members of the OpenAI “red team” who took to the stage
As mentioned by OpenAI, the “red team” system has existed for a long time, and many members have already received public interviews from the media.
In a revealing article in the Financial Times in April this year, Andrew White, a professor of chemical engineering at the University of Rochester in the United States, said that he participated in the "red team" test of GPT-4, and his attack goal was to use ChatGPT to create a new neurotoxin. White said he used the "plug-in" function to feed chemistry papers and chemical manufacturer catalogs to the big model, and the big model even found him a place where he could produce his hypothetical toxin.
White said that large models could eventually help many people do chemical experiments faster and more accurately, but there is also a serious danger that some people will use AI to do some dangerous experiments.
After receiving feedback from White, OpenAI made changes to the system to ensure that it would not respond to similar questions. According to the media, there are about 10 "red team" members with different backgrounds - lawyers, teachers, risk control personnel, security researchers, etc. Most of them are located in the United States and Europe. They revealed that in the months before GPT-4 was released, they spent roughly 10-40 hours testing the model, and were paid about $100 per hour.