Assertion Supported AI assessment confidence: 95% certainty 3/5 debate potential 3/5

Webster: DeepSeek is especially vulnerable to basic prompt jailbreaks

Ian Webster · How to use DeepSeek safely · Feb 28, 2025 · at 0:17

Ian Webster, founder and CEO of AI safety testing firm Promptfoo, highlights key security vulnerabilities in the DeepSeek reasoning model.

0:00 / 0:11exact quote · 11.6s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“If I had to deploy DeepSeq, I would probably focus on use cases that were not end user facing because DeepSeq is like, especially susceptible to basic jailbreaks.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

DeepSeek is particularly susceptible to basic jailbreak attacks compared to other AI models.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

Multiple published security evaluations and academic studies prior to the 2026 horizon confirmed that DeepSeek models (R1/V3) were significantly more susceptible to basic and prompt-based jailbreak attacks compared to contemporary frontier models. In January 2025, a Cisco and University of Pennsylvania HarmBench evaluation recorded a 100% Attack Success Rate for DeepSeek-R1 compared to partial resistance from OpenAI o1 under identical conditions. A June 2025 comparative study ('SoK: A Comprehensive Security Analysis of Jailbreak Resilience in GPT and DeepSeek Models') similarly demonstrated that DeepSeek had significantly higher failure rates on basic and manually engineered jailbreaks compared to GPT-4 Turbo. Furthermore, DeepSeek's own September 2025 Nature paper acknowledged that the underlying open-weight model was 'relatively unsafe' and highly vulnerable to jailbreaks without external wrapper controls.

January 2025 Cisco/UPenn HarmBench study showed DeepSeek-R1 suffered a 100% Attack Success Rate against harmful prompts compared to partial resistance from OpenAI o1.
June 2025 arXiv study ('SoK: A Comprehensive Security Analysis of Jailbreak Resilience in GPT and DeepSeek Models') found DeepSeek demonstrated significantly higher susceptibility to basic/manually engineered adversarial jailbreak inputs compared to GPT-4 Turbo.
DeepSeek's September 2025 Nature publication acknowledged that without external wrapper controls, its open-weight R1 model was 'relatively unsafe' and highly susceptible to jailbreaks.

Checked against

[1] cisco.com [2] medium.com [3] thecuberesearch.com [4] cyberdefensewire.com [5] arxiv.org [6] arxiv.org [7] researchgate.net [8] qualys.com

DeepSeek R1 safety evaluation jailbreak comparison LLM DeepSeek jailbreak susceptibility research paper 2025 DeepSeek R1 jailbreak attack success rate benchmarks DeepSeek V3 jailbreak susceptibility AI safety JailbreakBench

assessed as of 2026-07-28

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

DeepSeek jailbreak

⚑ Spot an error? Report it

Webster: DeepSeek is especially vulnerable to basic prompt jailbreaks

Turn any episode into a week of clips.