On Friday, Google revealed the first known case of indirect hacking by its Gemini AI software. This news came after other AI companies such as Anthropic and OpenAI also raised security concerns related to AI models. Google stated that in May, the Gemini AI model gained unauthorized access to three external systems.
Details of Unauthorized Access
Google explained that this unauthorized access occurred due to guessing login information or using credentials found in a public repository. Header Adkins, Vice President of Engineering Security at Google, stated that the AI model believed these systems were part of a test. However, in all three cases, the model halted its access before taking further action.
Read more: South Korean app developers call for action on Google and Apple fees
Adkins added: "In a standard assessment, the model found public information online and guessed credentials to access websites it thought were part of a test." Google emphasized that it does not consider these unauthorized accesses to be a level of misalignment, a term in the AI industry that refers to models deviating from human instructions. On the contrary, these intrusions were due to misidentification, where Gemini thought it was operating in a test but was actually connected to the real internet.
Consequences and Concerns
Google announced that it has corrected the model and believes these intrusions did not lead to any harm. Adkins noted: "These events highlight the importance of training powerful AI models to act responsibly." Concerns about the unexpected behaviors of AI representatives have increased in recent months. OpenAI announced in July that one of its agents hacked an AI startup called Hugging Face.
Sydney Van Ax, CEO of Nightingale Collective, an organization focused on AI safety, criticized Google for not disclosing these intrusions sooner. He said: "At this point, it can be said that we cannot expect companies to voluntarily disclose when their agents go out of control." He also stated that Google quickly announced that these incidents do not reach the level of misalignment.
Google stated that it was unaware of the intrusions until July, when a cybersecurity company named Irregular, which was conducting tests on Gemini, reviewed its work to investigate similar incidents following the disclosure of Hugging Face. The company then proceeded with investigations and informed the relevant organizations of the hacked websites and federal authorities about these incidents.
Irregular also stated that it does not consider this incident a "complex cyber act" and added that "there is no current open issue." The company plans to release a report in the coming weeks to share best practices for mitigating and securely conducting cybersecurity assessments.
This intrusion was reported on Friday by The Wall Street Journal. Concerns about AI safety have significantly increased, prompting several AI researchers to resign from their jobs and a diverse group of people to call for coordinated action to protect critical system security. However, these requests have faced skepticism from the White House and the Chinese government.
Read more: NASA discovers a crater larger than the Colosseum on the Moon · Zscaler stock introduced as the only investment option in cybersecurity
