EventPast event

Lasso Benchmark Compares Jev with Prompt Injection Classifiers

1 reports1 reporting sourcesLatest reportOct 14, 2026

Start here

AI overview

Lasso Security tested prompt injection classifiers against 3,669 public benchmark attacks. At a 0.25% false positive rate threshold, its proprietary classifier recalls 95.4%, TypeSafe's Jev 76.2%, and PIGuard and fine-tuned Laya about 30%. Jev recalls 99.8% on explicit override-style attacks, but only 63.4% on 2,383 injections with no override wording.

Generated by AI from reports · Updated 29 days ago

Latest updateOct 14, 2026

Lasso vs. Jev: Can Jev really replace prompt injection classifiers?

Related research

Papers whose work the reports mention

The reports mention LLMail-InjectLLMail-Inject: A Dataset from a Realistic Adaptive Prompt Injection Challenge
“We ran every system against 3,669 attacks from eight public benchmarks: InjecAgent, LLMail-Inject, Tensor Trust, AgentDojo, AgentDyn, BIPIA, WildChat and WildGuard.”
Lasso vs. Jev: Can Jev really replace prompt injection classifiers?
Releases the LLMail-Inject dataset to evaluate prompt injection defenses of email assistants using realistic adaptive human attacks
arXiv · 2025-06-11 · Deep reading
The reports mention InjecAgentInjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents
“We ran every system against 3,669 attacks from eight public benchmarks: InjecAgent, LLMail-Inject, Tensor Trust, AgentDojo, AgentDyn, BIPIA, WildChat and WildGuard.”
Lasso vs. Jev: Can Jev really replace prompt injection classifiers?
Proposes the InjecAgent benchmark to evaluate the vulnerability of tool-integrated agents to indirect prompt injection
arXiv · 2024-03-05 · Deep reading

Progress timeline

Oct 14

Time not given
Lasso Security
Lasso vs. Jev: Can Jev really replace prompt injection classifiers?
Lasso tested prompt injection classifiers on 3,669 public benchmark attacks: at a 0.25% false positive rate, its own classifier recalls 95.4%, TypeSafe's Jev 76.2%, and PIGuard and fine-tuned Laya about 30%. Jev recalls 99.8% on explicit override attacks but only 63.4% on 2,383 injections without override wording.