OpenAI's artificial intelligence (AI) agents may have invaded or caused negative impacts on the systems of over 100 organizations, as reported by the company on Wednesday night (30th).
OpenAI, responsible for ChatGPT, stated it has notified over 100 third-party organizations about agent activities considered “misaligned”. The disclosure expands the known scope of actions by agents that exhibited unexpected behaviors and raises new questions about the control that AI companies can maintain over their latest models during testing and evaluation phases.
The identified cases present different levels of severity. Among them are attempts to induce websites to execute unexpected commands and actions to bypass certain security controls without authorization.
This does not necessarily mean that a system has been effectively compromised.
The disclosure raises new questions about the control that AI companies can maintain over their latest models during testing and evaluation phases – Credit: Primakov/Shutterstock
Read more:
- OpenAI is negotiating a round of R$15.61 billion and may reach a valuation of R$7.3 trillion
- The US is investigating OpenAI and Anthropic after incidents involving AI agents
- OpenAI accuses Chinese Moonshot of distilling its AIs
OpenAI wants to share its findings
- According to OpenAI, the notifications are sent privately to the affected organizations;
- The goal is to provide companies with the necessary information to investigate and resolve potential security issues or other technical concerns;
- At the same time, the company stated that it intends to publicly disclose its findings on model behaviors and new types of vulnerabilities in security mechanisms;
- The intention, according to the company, is to allow AI and security research communities to use this information to improve system protection.
The disclosure occurs after independent researchers identified a series of incidents involving AI agents that exhibited behaviors considered unexpected in cybersecurity-related situations.
O The Washington Post reported on Wednesday (30) that AI agents with behavior similar to that observed in OpenAI's systems had attempted to infiltrate Canadian government websites.
Rodrigo Mozelli is a journalist graduated from the Universidade Metodista de São Paulo (UMESP) and, currently, is a writer at Olhar Digital.
