Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Re-approve the four unsloth-zoo findings that 2026.9.5 reopened (#11240)

Daniel Han16408b25fdef
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 4aa6d14c457..60f86826825 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -1953,6 +1953,42 @@
"evidence": "Obfusc: L4855: _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L158: mx.eval(model.parameters()) | L192: mx.eval([v for k, v in tree_flatten(casted) if is_mlx_norm_parameter_path(k)]) | L1254: mx.eval(model.parameters()) | L1255: mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L1322: mx.eval(model.parameters()) | L1325: mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L2711: mx.eval(altup.correct(predictions, activated)) | L3210: mx.eval(baseline, perturbed) | L4503: model.eval() | L5242: mx.eval(model.parameters()) | L5357: mx.eval(module.weight) | L8769: lambda: mx.eval(model.parameters()), | L8829: lambda: mx.eval(model.parameters()), | L8985: mx.eval(model.parameters())",
"evidence_hash": "cf9c8ed7bf9c33f83b6280d7d2d60b072262a04ce3e886a1f38c4003fbef01f4",
"file_sha256": "7a69b550899a2d782bb460e3d882a0840b1e2f8e7b50f43d6a9468ae808aad59"
+ },
+ {
+ "package": "unsloth-zoo",
+ "file": "unsloth_zoo/vision_utils.py",
+ "check": "Harvests environment variables/secrets AND makes network calls",
+ "severity": "CRITICAL",
+ "evidence": "Env: L151: return os.environ.get(UNSLOTH_ALLOW_PRIVATE_URL_FETCH_VAR, \"0\") == \"1\"\nNetwork: L109: from urllib.request import url2pathname | L478: session = requests.Session()",
+ "evidence_hash": "3bbd70075adbff1b2cb50de320002a0dce6324653a8727ec60c42e0943dccdac",
+ "file_sha256": "2fc55291dde4fde95a86b3386e6608db4d990af6d239471a3e46830798b714b3"
+ },
+ {
+ "package": "unsloth-zoo",
+ "file": "unsloth_zoo/vision_utils.py",
+ "check": "Accesses cloud metadata/IMDS AND makes network calls",
+ "severity": "CRITICAL",
+ "evidence": "IMDS: L144: \"metadata.google.internal\", \"metadata.goog\", \"metadata\",\nNetwork: L109: from urllib.request import url2pathname | L478: session = requests.Session()",
+ "evidence_hash": "972786b837f20621c6615cf1c38a0b84abe925d302d7317679bc0272643c0793",
+ "file_sha256": "2fc55291dde4fde95a86b3386e6608db4d990af6d239471a3e46830798b714b3"
+ },
+ {
+ "package": "unsloth-zoo",
+ "file": "unsloth_zoo/compiler.py",
+ "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+ "severity": "HIGH",
+ "evidence": "Obfusc: L1519: _mod = __import__(model_location, fromlist=items) | L5916: f\" {chr(92)}{chr(92)} /| Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL5917: f\"O^O/ {chr(92)}_/ {chr(92)} Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL5918: f\"{chr(92)} / Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL5919: f' \"-____-\" Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L5917: f\"O^O/ {chr(92)}_/ {chr(92)} Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL5918: f\"{chr(92)} / Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL5919: f' \"-____-\" Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L5916: f\" {chr(92)}{chr(92)} /| Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL5917: f\"O^O/ {chr(92)}_/ {chr(92)} Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL5918: f\"{chr(92)} / Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L1448: _modeling_file = eval(model_location) | L2154: f = eval(f\"{model_location}.{module}\") | L2495: exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L2496: try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L3978: exec(f\"import {parent}\", locals(), globals()) | L4092: dir(eval(parent)), | L4096: exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L4170: module = eval(f\"modeling_file.{module}\") | L4197: inner_class = eval(f\"modeling_file.{inner_class}\") | L4327: exec(f\"from timm.layers.norm_act import {norm}\") | L4335: forward = eval(norm).forward | L4341: exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L4358: exec(f\"from timm.models._efficientnet_blocks import {block}\") | L4366: forward = eval(block).forward | L4372: exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L4612: exec(\nL4613: f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4614: globals(),\nL4615: locals(),\nL4616: ) | L4618: exec(f\"{model_location}.nn.{module}.forward = forward\", globals(), locals()) | L4622: exec(\nL4623: f\"combined_module.torch.nn.{module}.forward = forward\",\nL4624: globals(),\nL4625: locals(),\nL4626: ) | L4628: exec(f\"combined_module.nn.{module}.forward = forward\", globals(), locals()) | L4705: source = eval(f\"{model_location}.torch\") | L4712: function = eval(f\"source.nn.{module}\") | L4906: exec(\"import transformers\", globals()) | L4919: exec(\nL4920: \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL4921: ) | L4923: exec(\nL4924: \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL4925: globals(),\nL4926: locals(),\nL4927: ) | L5161: source = eval(f\"modeling_file.{module}\") | L5175: source = eval(f\"modeling_file.{module}\") | L5276: source = eval(f\"modeling_file.{module}\") | L5314: source = eval(f\"{model_location}.{module}\") | L5394: source = eval(f\"{model_location}.{module}\") | L5442: source = eval(f\"{model_location}.{module}\") | L5679: source = eval(f\"{model_location}.{module}\") | L5690: exec(\nL5691: f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL5692: globals(),\nL5693: ) | L5756: source = eval(f\"{model_location}.{module}\") | L5797: module_cls = eval(f\"{model_location}.{module}\") | L5834: module_cls = eval(f\"{model_location}.{module}\") | L5901: exec(\nL5902: \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL5903: globals(),\nL5904: ) | L5966: exec(inner_training_loop, globals()) | L5975: function = eval(f\"{model_location}.{module}\") | L6062: function = eval(f\"{model_location}.{module}\") | L6216: exec(\nL6217: f\"{model_location}.{module} = combined_module.{module}\",\nL6218: globals(),\nL6219: locals(),\nL6220: ) | L6230: check_dicts = dir(eval(f\"{model_location}\")) | L6232: item = eval(f\"{model_location}.{check}\") | L6242: exec(\nL6243: f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL6244: globals(),\nL6245: locals(),\nL6246: )",
+ "evidence_hash": "bdfae4c24572c2a86fa6a2dede720ebb0723746963e43c2cd42701de5fc48fdb",
+ "file_sha256": "48254dd0bcbc23743b7a81f177bb1e7acda186c809063dc7f3ad0f262b9e7bdb"
+ },
+ {
+ "package": "unsloth-zoo",
+ "file": "unsloth_zoo/mlx/loader.py",
+ "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+ "severity": "HIGH",
+ "evidence": "Obfusc: L4855: _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L158: mx.eval(model.parameters()) | L192: mx.eval([v for k, v in tree_flatten(casted) if is_mlx_norm_parameter_path(k)]) | L1254: mx.eval(model.parameters()) | L1255: mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L1322: mx.eval(model.parameters()) | L1325: mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L2711: mx.eval(altup.correct(predictions, activated)) | L3210: mx.eval(baseline, perturbed) | L4503: model.eval() | L5242: mx.eval(model.parameters()) | L5357: mx.eval(module.weight) | L8769: lambda: mx.eval(model.parameters()), | L8829: lambda: mx.eval(model.parameters()), | L8985: mx.eval(model.parameters())",
+ "evidence_hash": "cf9c8ed7bf9c33f83b6280d7d2d60b072262a04ce3e886a1f38c4003fbef01f4",
+ "file_sha256": "a865c6b908885b4a7a7df38dc6960c645f1b60b011b720d741e02092ee239903"
}
]
}