experiment-bridge

Pass

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill reads from refine-logs/EXPERIMENT_PLAN.md, refine-logs/EXPERIMENT_TRACKER.md, and refine-logs/FINAL_PROPOSAL.md to extract execution commands and parameters. These files function as ingestion points for untrusted data that drives code generation and execution behaviors. Boundary markers are absent for these file reads. Capability inventory includes broad shell access and file-writing. No explicit sanitization or escaping of extracted hyperparameters is present.
  • [EXTERNAL_DOWNLOADS]: The workflow supports a BASE_REPO parameter that triggers a git clone from a remote URL. This enables the skill to download and then execute external code at runtime, which is then used as a base for model training and evaluation on the host machine.
  • [PROMPT_INJECTION]: Phase 2.5 contains a complex prompt for an automated reviewer (GPT-6-Astra) that instructs the model to ignore specific security risks such as local user threats and over-defense. These instructions deliberately attempt to lower the assistant's safety scrutiny for the generated experiment code.
  • [COMMAND_EXECUTION]: The skill's primary workflow involves the automated generation and execution of Python and Bash scripts on the host system or connected GPU instances (vast.ai, Modal). This capability is used to run training loops and evaluation suites directly.
  • [DYNAMIC_EXECUTION]: The skill creates executable training and evaluation scripts at runtime based on the analysis of experiment plans and method proposals. These generated scripts are executed to verify the sanity and main results of the proposed research.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 13, 2026, 02:26 AM