Cisco Exposes Critical Flaw in Enterprise LLMs
At the Black Hat 2025 cybersecurity conference in Las Vegas, Cisco unveiled a newly discovered method for bypassing protections in large language models (LLMs) — raising serious concerns for organizations using AI in sensitive environments.
The method, called “instructional decomposition”, enables attackers to extract confidential data by cleverly breaking down queries. Instead of directly requesting protected content, the attacker first asks for a high-level summary. Once a general context is set, they proceed to request small fragments of the desired information. These pieces can then be reassembled to reconstruct the entire original content, all without triggering existing AI guardrails.
During their demonstration, Cisco researchers used this technique to successfully recover full articles from The New York Times, completely evading safety filters.
A Growing Threat to AI Systems
New data from IBM shows that 13% of security breaches now involve AI systems, most through jailbreak exploits. Even more alarming: 97% of affected companies had insufficient access controls in place for their AI models or chatbots.
The potential damage is significant. Experts warn that attackers could access intellectual property, classified business data, or personal information, especially in companies using LLMs trained on internal or confidential datasets.
What Can Be Done?
Security professionals recommend a combination of:
-
Rigorous prompt monitoring
-
Restricting access to AI tools
-
Developing smarter detection systems for prompt manipulation
Cisco notes that while it’s nearly impossible to block every type of jailbreak, proactive measures can dramatically reduce exposure and improve LLM security posture.
Moving The Sea With Us!
Contact us today: +351 265 544 370 or go to Contacts Page
Email: sales@partyard.eu
#PartYard #CyberSecurity #Cisco #LLMSecurity #AIVulnerabilities #BlackHat2025 #DataProtection #InstructionalDecomposition #EnterpriseAI #AIJailbreak #SensitiveData #AISecurity #TechNews #Infosec #AICompliance
