Jailbreak Guardrails Benchmark
Benchmarking Jailbreak Guardrails for Embodied Agents
A framework evaluating 6 guardrails on embodied agents against jailbreaks and safe instructions, measuring defense, usability, and efficiency.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info jailbreak-guardrails-benchmark
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-4o | 40.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| DeepSeek V3.2 | 40.3 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
The benchmark evaluates six representative guardrails integrated into embodied agent pipelines at perception, planning, or control stages. It uses template-based and automated jailbreak attacks alongside safe instructions to assess system-level performance.
Scoring focuses on three dimensions: defense effectiveness (bypass/hazard rates), usability (false-positive/task completion rates), and efficiency (latency overhead). Results reveal trade-offs among these dimensions, showing no single guardrail dominates across all settings.