第 12 份独立评测：注入鲁棒性系统测试《Can Jev be jailbroken?》（tanh.xyz，14,495 次记录调用，原始记录公开）
Research note testing whether Jev's typed answers can be manipulated by prompt injection, across 14,495 logged calls with raw logs.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
"Can Jev be jailbroken?" is a research note on tanh.xyz, listed as the 12th independent evaluation of Jev. It systematically tests injection robustness: whether text placed in the state can push Jev's typed answers away from what they should be.
The listing reports 14,495 logged calls and says the raw records are public. It matters to anyone using a decision model as a guardrail, since a gate that can be talked out of its answer is weak.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Can injected instructions in the state change Jev's typed answers?(inferred) | — |
How it works
- Build prompts with injected manipulation attempts in the state
- Send them to Jev with typed questions
- Log every call and compare answers with and without injection
- Publish the raw records
Strengths
- Large, logged test set (14,495 calls, per the listing)
- Raw records reportedly public for reanalysis
Caveats
- Details beyond the title come from a truncated list summary
- Results apply to the model version tested
At a glance
- Type
- Article
- Status
- research
Use cases and integrations
How curated lists describe it
The 12th independent evaluation: systematic injection-robustness testing, “Can…
wh000wh000/awesome-jev-live
Frequently asked questions
What does it test?
Whether prompt-injection text in the input can manipulate Jev's typed decisions.
How big is the test?
The listing cites 14,495 logged calls with raw records published.
Where is it?
https://tanh.xyz/research-notes/jev-prompt-manipulation
Related projects
Test harness for any Jev-API decision model: accuracy, calibration and robustness probes for reordering, injection and padding.
ServicebetaPythonPaper code where Jev screens agent tool output for relevance and prompt injection before the agent reads it.
ProjectresearchPythonResearch code showing that short, natural-looking context additions can flip decision models like Jev to a chosen wrong answer.
ProjectresearchPythonBenchmark of Jev versus Claude Haiku 4.5 as input safety classifiers on jailbreak, evasion and benign pseudo-harmful prompts.
BenchmarkresearchPythonBlind security benchmarks for Jev on public corpora: prompt-injection detection (96.5% with context) and vulnerable-code detection.
BenchmarkresearchGoPaper testing Jev, Laya, Decider and Nimble against classifiers and LLM judges on prompt injection and risk screening.
Paperresearch