I don't know if Anthropic, OpenAI, or Google have really spent a lot of time and effort on security around their models. I know they talk about it (Anthropic, OpenAI, Google), but in practice, I'm not sure the researchers and testers are being as careful as the marketing and management of those companies want you to think.
Gemini broke into three companies during a security test. Claude escaped a test environment and breached three organizations. OpenAI had breaches of Hugging Face and other sites, so perhaps this is the most powerful model, as it exceeded the "rule of three." In all these cases, the agents weren't doing anything amazing. These were acting the same as most malicious or bored human hackers who might find or guess credentials and take advantage of them. In some sense, this points to the generally poor state of security in the world.
Podden och tillhörande omslagsbild på den här sidan tillhör
Steve Jones. Innehållet i podden är skapat av Steve Jones och inte av,
eller tillsammans med, Poddtoppen.
