Major LLMs Face ISC Security Risk as Research Shows Legitimate Tasks Can Trigger Dangerous Outputs

Researchers from Deakin University, Fudan University’s Institute of Trustworthy Embodied AI, City University of Hong Kong and other institutions have dubbed the risk “Internal Safety Collapse,” or ISC. When LLMs encounter ostensibly legitimate professional workflows that require sensitive data to complete, they may bypass existing alignment and refusal safeguards. The findings show that risks arise not only from malicious prompts but also within workflows involving AI agents and dual-use tools.

The team released its paper and ISC-Bench on March 4, 2026, covering 53 scenarios across eight professional domains. Using JailbreakBench to test GPT-5.2, Claude Sonnet 4.5, Gemini 3 Pro and Grok 4.1, researchers found an average safety failure rate of 95.3% under the worst-performing configuration. The API cost per attack objective was as low as $0.002, and none of the models proactively refused any of the task patterns tested.