Back to Paper
cs.CR
Local ID: 2604.19461v2
AI Summary: gemma4:e4b
Involuntary In-Context Learning: Exploiting Few-Shot Pattern Completion to Bypass Safety Alignment in GPT-5.4
By Alex Polyakov, Daniel Kuznetsov
Revision History Timeline
v14/21/2026
4/21/2026
No submitter comment provided.
v26/3/2026
6/3/2026
No submitter comment provided.
★ Version indexed in ExplorerComparing v1 vs v2
Green = Added • Red = Removed
Title Comparison
Involuntary In-Context Learning: Exploiting Few-Shot Pattern Completion to Bypass Safety Alignment in GPT-5.4
Authors Comparison
No author changes.
v1 Comment
No comment for this version.
v2 Comment
No comment for this version.
Abstract Word Diff
Safety alignment in large language models relies on behavioral training that can be overridden when sufficiently strong in-context patterns compete with learned refusal behaviors. We introduce Involuntary In-Context Learning (IICL), an attack class that uses abstract operator framing with few-shot examples to force pattern completion that overrides safety training. Through 3479 probes across 10 OpenAI models, we identify the attack's effective components through a seven-experiment ablation study. Key findings: (1)~semantic operator naming achieves 100\,\%100% bypass rate (50/50, $p < 0.001$); (2)~the attack requires abstract framing, since identical examples in direct question-and-answer format yield 0\,\%;0%; (3)~example ordering matters strongly (interleaved: 76\,\%,76%, harmful-first: 6\,\%);6%); (4)~temperature has no meaningful effect (46--56\,\%(46-56% across 0.0--1.0). On the HarmBench benchmark, IICL achieves 24.0\,\%24.0% bypass $[18.6\%,$[18.6%, 30.4\%]$30.4%]$ against GPT-5.4 with detailed 619-word responses, compared to 0.0\,\%0.0% for direct queries.