Fix Pickle Vulnerability Detection Gap

[ready-research]

PR Description: Fix Pickle Vulnerability Detection Gap

Summary

Fixes critical detection gaps in pickle vulnerability scanning by improving pickle parsing engine and expanding unsafe globals coverage.

Fixes: #311

Problem

Modelscan was missing significant number of malicious pickle files:

• Before fix: 22/60 detections (33.7% detection rate)
• Missing: 38 malicious files including all GHSA vulnerabilities
• Impact: Major security exposure for applications using pickle files

Solution

Enhanced Pickle Parsing Engine

• Incremental opcode parsing with deferred error handling for broken streams
• Expanded STACK_GLOBAL support for classic string opcodes (STRING, BINSTRING, SHORT_BINSTRING)
• Improved error resilience to extract globals from partially corrupt files
• Better unknown detection for non-string opcode values

Comprehensive Unsafe Globals Database

Added detection patterns for:

• PyTorch GHSA vulnerabilities: torch._dynamo.guards.GuardBuilder.get, torch.utils.bottleneck.__main__.*, torch.fx.experimental.symbolic_shapes.*, etc.
• System execution risks: commands, ssl, timeit, ensurepip._run_pip
• Dynamic code execution: code.InteractiveInterpreter.runcode, cProfile.run, profile.*
• Development tools: idlelib.*, lib2to3.*, pydoc.pipepager

Robust Archive Handling

• Per-entry error handling in zip files
• Continue scanning after encountering corrupt entries
• Clear skip logging with BAD_ZIP categorization

Test Results

• Before: 22/60 detections (33.7%)
• After: 60/60 detections (100.0%)
• Improvement: +38 additional detections
• Fixed: All GHSA vulnerabilities now detected

Files Modified

• modelscan/tools/picklescanner.py - Core parsing improvements
• modelscan/settings.py - Expanded unsafe globals database
• modelscan/modelscan.py - Enhanced zip error handling

Validation

Test against comprehensive pickle vulnerability suite: Sent in email

This patch transforms modelscan from detecting ~36% to ~100% of malicious pickle files, significantly improving security coverage.

Note: Inspiration from picklescan project.

[ready-research]

@swashko @iamfaisalkhan @theinfosecguy Please review this PR.