'

Google’s Gemini AI Vulnerability Lets Attackers Gain Control Over Users’ Queries

Researchers at HiddenLayer have unveiled a series of vulnerabilities within Google’s Gemini AI that could allow attackers to manipulate user queries and control the output of the Large Language Models (LLMs).

This revelation has raised concerns over the security and integrity of AI-driven content generation and its implications for misinformation spread and data privacy.

The Gemini suite, Google’s latest foray into the realm of LLMs, comprises three different model sizes: Nano, Pro, and Ultra.

Despite its recent removal from service due to the generation of politically biased content, the vulnerabilities identified by HiddenLayer expose a new dimension of threats that malicious actors could exploit.

Document

Integrate ANY.RUN in your company for Effective Malware Analysis

Are you from SOC and DFIR teams? – Join With 400,000 independent Researchers

Malware analysis can be fast and simple. Just let us show you the way to:

  • Interact with malware safely
  • Set up virtual machine in Linux and all Windows OS versions
  • Work in a team
  • Get detailed reports with maximum data
  • If you want to test all these features now with completely free access to the sandbox:

The Vulnerabilities Explained

The research conducted by HiddenLayer highlights several critical issues within the Gemini models, including:

  • LLM Prompt Leakage: This vulnerability could allow attackers to access sensitive data or system prompts, posing a significant risk to data privacy.
Google’s Gemini AI Vulnerability Lets Attackers Gain Control Over Users’ Queries
LLM Prompt Leakage
  • Jailbreaks: By bypassing the models’ safeguards, attackers can manipulate the AI to generate misinformation, especially concerning sensitive topics like elections.
If we ask Gemini Pro to generate our article conventionally, we unfortunately get this response:
If we ask Gemini Pro to generate our article conventionally, we, unfortunately, get this response
  • Indirect Injections: Attackers can indirectly manipulate the model’s output through delayed payloads injected via platforms like Google Drive, further complicating the detection and mitigation of such threats.
We can input a few different variants of uncommon tokens to get a reset response
We can input a few different variants of uncommon tokens to get a reset response

Implications and Concerns

The vulnerabilities within Google’s Gemini AI have far-reaching implications, affecting a wide range of users:

  • General Public: The potential for generating misinformation directly threatens the public, undermining trust in AI-generated content.
  • Companies: Businesses utilizing the Gemini API for content generation may be at risk of data leakage, compromising sensitive corporate information.
  • Governments: The spread of misinformation about geopolitical events could have serious implications for national security and public policy.

Google’s Response and Future Steps

As of the publication of this article, Google has yet to issue a formal response to the findings.

The tech giant previously removed the Gemini suite from service due to concerns over biased content generation. Still, the new vulnerabilities underscore the need for more robust security measures and ethical guidelines in the development and deployment of AI technologies.

The discovery of vulnerabilities within Google’s Gemini AI is a stark reminder of the potential risks associated with LLMs and AI-driven content generation.

As AI continues to evolve and integrate into various aspects of daily life, ensuring the security and integrity of these technologies becomes paramount.

The findings from HiddenLayer highlight the need for ongoing vigilance and prompt a broader discussion of AI’s ethical implications and the measures needed to safeguard against misuse.

You can block malware, including Trojans, ransomware, spyware, rootkits, worms, and zero-day exploits, with Perimeter81 malware protection. All are incredibly harmful, can wreak havoc, and damage your network.

Stay updated on Cybersecurity news, Whitepapers, and Infographics. Follow us on LinkedIn & Twitter

The post Google’s Gemini AI Vulnerability Lets Attackers Gain Control Over Users’ Queries appeared first on GBHackers on Security | #1 Globally Trusted Cyber Security News Platform.


Go to Source
Author: Divya