sglang-sota-performance

Pass

Audited by Gen Agent Trust Hub on May 3, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes untrusted user-provided models and datasets.
  • Ingestion points: User-supplied model_id, local checkpoint paths, and workload datasets used for benchmarking and profiling.
  • Boundary markers: Absent; there are no specific instructions to use delimiters or ignore embedded instructions within the ingested data.
  • Capability inventory: The skill can execute various benchmarking and profiling shell commands and perform file-write operations to patch local source code.
  • Sanitization: Absent; no input validation or escaping mechanisms are described before data is used in commands or logic.
  • [COMMAND_EXECUTION]: The skill requires the agent to execute shell-based tools for benchmarking and serving (e.g., trtllm-serve) and profiling. It also involves the dynamic generation and application of source code patches to the SGLang framework based on performance analysis.
Audit Metadata
Risk Level
SAFE
Analyzed
May 3, 2026, 03:23 AM