For Gemini Ultra, we’re currently completing extensive trust and safety checks, including red-teaming by trusted external parties, and further refining the model using fine-tuning and reinforcement learning from human feedback (RLHF) before making it broadly available.
Thanks.
Thanks for the term, I’ve never seen one before.
If I have well understood, red teams are hackers hired to do the same things legally, so to stress-test the security capacity of an organization. Funny thing, they attack the blue teams, that instead operate to defend the organization.
Both teams not necessarily share information, this is why there are purple teams that have the purpose to help facilitate communication between the two, and improve the system.
Fan fact, between 2023-2025 the European parliament have required to all major European banks to go through an increasing stress-test for their systems. Banks are required to be able to guarantee an overall performance of all critical operations, during an attack. Plus, of course, protect their money!
But I don’t know much more about it, I have just quickly read something about.
1-sentence version:
A red team is a team of hackers, penetration testers, or security personnel that a company employs or contracts to test the company’s security by trying to gain a specified level of access to a specific system over a longer period (weeks or months).
A bit of unpacking:
Essentially, red teams use all methods available to hackers to try to compromise the system to reach their stated goal. A traditional penetration test (pentest) is on a shorter period (<2 weeks) and uses a limited scope and toolset to find common vulnerabilities. You can think of red team events as an “is it practicable to get to X on the network” test, whereas a pentest is more of an exercise to determine what’s easily exploitable using simple methods.
