Summary

Closes the security concern raised in the issue: an attacker-controlled overriding_architecture_config (inline dict or JSON file) could inject {"trust_remote_code": true}, which would be splatted directly into AutoModelForCausalLM.from_pretrained() / AutoModelForSeq2SeqLM.from_pretrained(), enabling arbitrary remote code execution.

Previously the fix was an inline _BLOCKED_KEYS block inside HuggingFaceTransformersDecoder.__init__. This PR extracts that logic into a reusable get_safe_config() helper and applies it everywhere overriding_architecture_config reaches from_pretrained.

Changes

espnet2/asr/decoder/hugging_face_transformers_decoder.py

• Removes inline _BLOCKED_KEYS / dict-comprehension logic from __init__
• Adds module-level _BLOCKED_CONFIG_KEYS = {"trust_remote_code"} constant
• Adds get_safe_config(config: dict) -> dict public helper that strips blocked keys and logs a warning for each one removed; handles None/non-dict inputs gracefully
• Replaces the inline block with self.overriding_architecture_config = get_safe_config(...)

espnet2/bin/asr_inference.py

• Imports get_safe_config from the decoder module
• Wraps decoder.overriding_architecture_config with get_safe_config() at both AutoModelForCausalLM.from_pretrained and AutoModelForSeq2SeqLM.from_pretrained callsites (defense-in-depth)