Instructional decomposition exposes AI guardrail flaws

The IBM 2025 Cost of a Data Breach Report shows 13% of all data breaches now involve company AI models or apps, with most incidents exploiting jailbreak techniques. A jailbreak breaks free from developer guardrails to extract original training data, and Cisco’s “instructional decomposition” demo at Black Hat revealed how context manipulation and incremental prompts can bypass those defenses. By first obtaining an oblique summary of copyrighted material and then requesting individual sentences without naming the full source, attackers can reconstruct verbatim content without triggering filters. Since jailbreaks are unlikely to be fully prevented, experts stress the need for stricter AI access controls, continuous monitoring, and robust governance.

Read more:

https://www.securityweek.com/ai-guardrails-under-fire-ciscos-jailbreak-demo-exposes-ai-weak-points/