Security Research into Chinese Open-Weight Models

Narrative

Independent technical bloggers were first movers on DeepSeek scrutiny. Simon Willison tested DeepSeek-R1 directly within days of release, probing prompt-injection resistance and noting that a chain-of-thought hack (intercepting the closing think-tag) could force the model to extend reasoning indefinitely, alongside his broader point that language models are "inherently gullible" as a technology, not a China-specific flaw. HiddenLayer's "DeepSh*t" post used its proprietary automated red-teaming and ShadowGenes model-genealogy tooling and manual assessment to conclude that R1 was vulnerable to jailbreak techniques, prompt injections, glitch tokens, and exploitation of its control tokens, making it less secure than other modern LLMs