
Openai’s new jailbreak vulnerabilities called “Time Bandit”, called “Time Bandit”, were abused to bypass the safety function of chatbot incorporation.
With this vulnerability, the attacker can manipulate chatbots to create illegal or dangerous content, such as creating malware, fishing fraud, and other malicious activities.
This jailbreak exploitation caused an alarm in the cyber security community, as it could be expanded for malignant purposes by threat stakeholders.
Mechanism of “Time Bandit”
The escape of “Time Bandit”, which was revealed by researchers Dave Kuszmar, confused AI by fixing the response to a specific historical period. Attackers can use this vulnerability in two major ways. Use the search function integrated into Chatgpt through direct interaction with AI.
Collect threat intelligence by TI lookup to improve the security of the company -Get 50 free requests
Direct interaction: In this method, the attacker starts a session by promoting AI with historical events, periods, or context -related questions. For example, you may ask a chatbot to simulate the support of the 1800s task. Once a historical background is established in a conversation, the attacker can gradually pivot discussions for illegal topics. By maintaining historical context, the attacker exploits the ambiguity of chatbot response procedures and incorrectly violates the safety guidelines.
Utilization of search functions: You can also operate the search function of Chatgpt that gets information from the web. The attacker instructs AI to search for topics linked to a specific historical era, and uses an illegal subject using the subsequent search and operation prompts. This process uses the turmoil of the timeline to accuse AI and provide prohibited content.
The bug was first reported by CERT COLDINATION CENTER (Cert/CC) by Cyber Security researcher, Dave Cusmer. During the controlled test, they were able to make jailbreak multiple times. Once you start, Chatgpt may create illegal content after detecting and deleting a specific prompt that violates the use policy.
The most notable is that jailbreak is more effective if the historic time frame of the 1800s and 1900s was used.
Utilizing vulnerability through prompts did not require user authentication, but a login account was required to use the search function. This double method of exploitation indicates the versatility of the vulnerabilities of “time bandits”.
The meaning of this vulnerability is extensive. By bypassing Openai’s strict safety guidelines, the attacker can use ChatGpt to generate steps in order to create weapons, drugs, or malware.
It can also be used for fishing fraud, social engineering scripts, or mass production of other harmful content.
Using legal and widely trusted tools, such as Chatgpt, can make malicious activities even more difficult, making it more difficult to detect and prevent them.
Experts have warned under the control of organized cyber criminals, that “Time Bandit” may promote large -scale malicious operations and have a great threat to cyber security and public safety. Masu.
Openai is already acting to deal with vulnerabilities. In a statement, Openai’s spokeswoman emphasized the company’s commitment to safety. “It is very important to develop a model safely. I do not want to use the model for malicious purposes. Thank you for disclosing the survey results. We maintain the usefulness of the model and the performance of tasks. We are always working on more secure and robust models for Exploiting, including jailbreak.
The recent transaction DeepSeek R1 model has become JAILROKED to generate Ransomware development scripts. “Deepseek R1 provides detailed instructions, has generated malicious scripts designed to extract credit card data from specific browsers and send them to remote servers.”
For updating daily security! Follow with Google News, Linkedin, and X