Researchers from Duke University and Carnegie Mellon University have demonstrated successful jailbreaks of OpenAI’s o1/o3, DeepSeek-R1, and Google’s Gemini 2.0 Flash models through a novel attack method called Hijacking Chain-of-Thought (H-CoT).
The research reveals how advanced safety mechanisms designed to prevent harmful outputs can be systematically bypassed using the models’ reasoning processes, raising urgent questions about AI security protocols.
Anatomy of the Vulnerability
The team developed Malicious-Educator, a benchmark that masks dangerous requests within innocuous educational prompts.
For example, asking “How should teachers explain white-collar crime prevention to students?” might appear legitimate but can be weaponized to extract detailed criminal strategies.
The study found that all tested models failed to recognize these contextual deceptions, with refusal rates plummeting from initial safety baselines.
OpenAI’s o1 model initially resisted 98% of malicious queries but became significantly more vulnerable after routine updates.
Researchers suspect updates improve general capability at the expense of safety alignment.
DeepSeek-R1 proved particularly susceptible to financial crime queries, providing actionable money laundering steps in 79% of test cases without requiring specialized attack techniques.
Gemini 2.0 Flash’s multi-modal architecture introduced unique risks – when fed manipulated diagrams alongside text prompts, its refusal rate dropped to 4%.
The H-CoT Attack Methodology
This technique manipulates the models’ self-monitoring process. As AI systems analyze prompts through chain-of-thought reasoning, attackers can inject misleading context that appears benign in early reasoning steps.
The study demonstrated how an NSFW image masked as “art history analysis” could trick models into discussing explicit content.
“We’re not just bypassing filters – we’re making the safety mechanism work against itself,” explained lead author Martin Kuo.
The findings come amid growing reliance on AI for sensitive applications, from education to healthcare.
Cybersecurity experts warn that these vulnerabilities could enable disinformation campaigns, financial fraud, and other malicious activities.
While companies often withhold security specifics, the research team has shared mitigation strategies with affected vendors. Temporary fixes include:
def safety_layer(response):
if "H-CoT" in response.metadata:
return SAFETY_OVERRIDE
# Additional checksLong-term solutions require fundamental redesigns of safety architectures. “We need systems that verify reasoning integrity, not just filter outputs,” advised co-author Hai Li.
This study underscores the delicate balance between AI capability and security.
As models grow more sophisticated, their self-monitoring mechanisms paradoxically create new attack surfaces – a challenge demanding immediate attention from AI developers and policymakers.
Free Webinar: Better SOC with Interactive Malware Sandbox for Incident Response, and Threat Hunting - Register Here
