SABER: Benchmarking Operational Safety of LLM Coding Agents

SABER is a benchmark for evaluating the operational safety of large language model coding agents in stateful project workspaces. The dataset includes task definitions, a sandbox runtime, and judging pipelines used for the research paper. It was created by sssr-lab and last updated on June 5, 2026.

Use Cases

• Benchmarking LLM coding agents on operational safety tasks based on the included task definitions.
• Reproducing baseline safety evaluations using the provided judging pipeline and utilities.
• Developing new safety tests for AI code generation within the provided sandbox runtime environment.

Strengths

• Includes a complete benchmark suite with task definitions, runtime, and judging tools as described.
• Provides utilities for reproducing baseline results from the associated research paper.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-06-05 13:17:28; freshness should be verified.

Provenance

Source

sssr-lab

Collection Method

Created as the code release for a research paper on LLM coding agent safety.

Freshness

Last updated 2026-06-05 13:17:28.