HiddenLayer researchers find TypeSafe’s Jev model more vulnerable to attacks than traditional LLMs
Security researchers report that TypeSafe’s streamlined Jev model lacks the defensive checks found in standard large language models, making it susceptible to injection.
Key points
- HiddenLayer researchers found TypeSafe’s Jev model is vulnerable to adversarial text attacks and prompt injection.
- The model’s streamlined design removes some checks and balances present in traditional large language models.
- TypeSafe did not respond to requests for comment regarding the security findings published Wednesday.
Researchers at HiddenLayer published a report on Wednesday stating that TypeSafe’s Jev model is more susceptible to certain attacks than traditional large language models. The security vendor argued that the model’s streamlined architecture creates vulnerabilities that allow adversarial text and prompt injection to disrupt outputs.
Kasimir Schulz, director of security research at HiddenLayer, co-authored the blog post detailing these findings. He explained that the accelerated approach used by Jev removes some of the protective checks found in standard LLMs. Consequently, the model can be tricked by attacks that current large language models typically detect and refuse.
Technical Vulnerabilities and Design Trade-offs
The report highlights that Jev is vulnerable to the TokenBreak attack out of the box, a technique that normally fails against larger LLMs. While simply inserting characters before important words did not yield significant results, the researchers noted that increasing perturbations improved attack success rates.
Schulz acknowledged that Jev offers faster response times and should be used for its speed advantages. However, he cautioned that the model lacks the full suite of protections seen in traditional systems. This means some attacks succeed against Jev that would have failed against other models.
Classification and Vendor Response
There is ongoing debate regarding whether Jev qualifies as a large language model. TypeSafe suggests it is not, while HiddenLayer researchers argue it appears to be one. TypeSafe has categorized Jev under a new class called System One, describing it as frontier models built for fast, structured decisions.
Jev is the first and only named product in the System One category. It is designed for machine-to-machine functioning, allowing it to perform tasks faster and use fewer tokens than traditional models. Despite these efficiency gains, the security report indicates that System One models appear vulnerable to specific attack vectors.
TypeSafe did not respond to multiple requests for comment regarding the HiddenLayer findings. The vendor has not issued a statement addressing the specific vulnerabilities or the classification debate raised by the security researchers in their Wednesday publication.
What should you do?
Organizations using Jev or similar System One models should review their input validation processes to mitigate prompt injection risks. Security teams should monitor model outputs for unexpected deviations or incorrect actions that could indicate successful adversarial attacks.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: TypeSafe
- Implement strict input sanitization to detect and block adversarial text patterns before they reach the model.
- Monitor model outputs for anomalies or unauthorized actions that may signal a successful prompt injection attempt.
- Restrict the model’s permissions to limit the potential impact of any incorrect or manipulated decisions.
- Evaluate the trade-offs between speed and security when deploying streamlined models in critical machine-to-machine workflows.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Is TypeSafe’s Jev model considered a large language model?
TypeSafe states that Jev is not an LLM, but HiddenLayer researchers argue that it appears to be one.
What specific attacks is Jev vulnerable to?
Jev is vulnerable to adversarial text attacks, prompt injection, and the TokenBreak attack technique.
Did TypeSafe respond to the security report?
No, TypeSafe did not respond to multiple requests for comment regarding the findings.
