InfoSecBulletin Cybersecurity for mankind
This week, multiple research teams showcased jailbreaks for popular AI models, including OpenAI’s ChatGPT, DeepSeek, and Alibaba’s Qwen.
After its launch, the open-source R1 model by Chinese company DeepSeek caught the attention of the cybersecurity industry. Experts found that jailbreak methods, previously patched in other AI models, still function against DeepSeek.
India’s Tata Electronics hit by cyber breach: Hacker target 630 GB record
Anthropic’s Mythos reportedly broke NSA classified systems in hours
OpenAI New Method “Deployment Simulation” Predicts AI Risks Before Deployment
AryStinger botnet infected thousands of D-Link routers globally
Hacker suspected of sending alerts across Brazil
CyberSentinel AI features 33 security tools like Nmap, SQLMap, and ZAP, utilizing Claude and GPT
Barracuda hosts Dhaka roundtable on cyber resilience
CISA Alerts Fortinet Users as FortiBleed Affects 86,644 FortiGate Devices
CISA: Splunk flaw under active exploit, patch by Sunday
Texas data breach exposes 3 million driver’s licenses
AI jailbreaking allows attackers to bypass safeguards designed to stop LLMs from producing harmful content. Security researchers have demonstrated that methods like prompt injection and model manipulation can overcome these protections.
Threat intelligence firm Kela found that DeepSeek is affected by Evil Jailbreak, where a chatbot is made to act as an evil confidant, and Leo, which allows the chatbot to take on an unrestricted persona. ChatGPT has fixed these vulnerabilities.
Palo Alto Networks’ Unit42 reported that DeepSeek is vulnerable to known AI jailbreak techniques.
The security firm successfully conducted the attack known as Deceptive Delight, which tricks generative AI models by embedding unsafe or restricted topics in benign narratives. This method was tested in the fall of 2024 against eight LLMs with an average success rate of 65%.
Palo Alto has successfully executed the Bad Likert Judge jailbreak, which asks the LLM to evaluate the harmfulness of responses using a Likert scale and generate examples that fit the scale.
Researchers discovered that DeepSeek is vulnerable to Crescendo, a jailbreak method that begins with innocuous dialogue and gradually shifts towards forbidden topics.
Alibaba has announced a new version of its Qwen AI model, claiming it outperforms the DeepSeek model.
Kela announced on Thursday that Alibaba’s new Qwen 2.5-VL model has vulnerabilities similar to those recently discovered in DeepSeek.
Researchers at a threat intelligence firm found that jailbreaks designed for DeepSeek also work on Qwen. They successfully tested an existing jailbreak called Grandma, which tricks the model into sharing dangerous information by having it role-play as a grandmother.
Kela found that Qwen 2.5-VL created content about developing ransomware and other malware.
“The ability of AI models to produce infostealer malware instructions raises serious concerns, as cybercriminals could leverage these capabilities to automate and enhance their attack methodologies,” Kela said.
Many jailbreak methods for ChatGPT have been fixed over the years, but researchers still discover new ways to bypass its protections.
CERT/CC reported that researcher Dave Kuszmar found a ChatGPT-4o jailbreak vulnerability called Time Bandit. This vulnerability allows users to ask the AI about specific historical events or instruct it to pretend to assist in such events.
“The jailbreak can be established in two ways, either through the Search function, or by prompting the AI directly,” CERT/CC explained in an advisory. “Once this historical timeframe has been established in the ChatGPT conversation, the attacker can exploit timeline confusion and procedural ambiguity in following prompts to circumvent the safety guidelines, resulting in ChatGPT generating illicit content. This information could be leveraged at scale by a motivated threat actor for malicious purposes.”
