Skip to content

Hide Navigation Hide TOC

Many-Shot Repetition Override via Extended Q&A Chain - ATR-2026-01005 (19becb45-7817-5ebd-9a60-8dac7d330c5d)

Detects many-shot prompt injection where the attacker builds an extended Q&A chain (5+ pairs) on benign topics (algebra, baking, Python syntax, system override definitions) to condition the model's compliance mode, then appends a malicious request. The defining artifact is a long Q/A: pattern sequence (≥3 Q:/A: pairs discernible in the input) combined with a sensitive terminal question. Separate from few-shot (ATR-2026-01004) which targets ≤3 pairs; many-shot relies on volume to overwhelm attention or context-window safety checks. Also detects the "Repeat X 10 times ... Q: [sensitive]" token-repeat conditioning variant.

Cluster A Galaxy A Cluster B Galaxy B Level
Many-Shot Repetition Override via Extended Q&A Chain - ATR-2026-01005 (19becb45-7817-5ebd-9a60-8dac7d330c5d) Agent Threat Rules LLM Prompt Injection (19cd2d12-66ff-487c-a05c-e058b027efc9) MITRE ATLAS Attack Pattern 1
Many-Shot Repetition Override via Extended Q&A Chain - ATR-2026-01005 (19becb45-7817-5ebd-9a60-8dac7d330c5d) Agent Threat Rules LLM Jailbreak (172427e3-9ecc-49a3-b628-96b824cc4131) MITRE ATLAS Attack Pattern 1