99 adversarial PE files: exploring malformed‑binary behaviour across major analysis tools

I’ve built a 99‑fixture adversarial PE corpus to explore how different tools behave when confronted with deliberately malformed but still loadable binaries.

Each fixture introduces one corruption pattern - no packers or multi‑anomaly noise, which allows for clean attribution of behaviour. The anomalies span:

• entrypoint redirection
• overlapping/invalid sections
• header inconsistencies
• directory OOB conditions
• TLS edge cases
• recursive/malformed resources
• Authenticode structural corruption
• entropy‑field manipulation

I tested 6 tools commonly used in exploit dev workflows:

• IOCX
• Ghidra
• Detect It Easy
• radare2
• PEview
• CFF Explorer

Behavioural patterns with exploit‑relevant implications:

• Literal parsers (r2, PEview) never crashed but provided no anomaly visibility
• Semantic parsers (CFF) “fixed” corruption, masking exploit‑useful inconsistencies
• Heuristic tools (DIE) ignored structure, blind to malformed metadata
• Reconstructive loaders (Ghidra) rewrote metadata, omitted fields, and crashed on entropy fixtures
• Hybrid literal‑semantic tools (IOCX) preserved raw bytes and surfaced anomalies explicitly

For exploit dev, malformed PE structures can act as:

• parser differentials
• crash primitives
• metadata confusion vectors
• loader‑model inconsistencies
• analysis‑evasion surfaces

This corpus maps those behaviours systematically.

Full write‑up (Part 1):

The Adversarial PE Analysis Series — Why PE Parsers Break

Corpus and fixture spec: https://github.com/iocx-dev/iocx

(fixtures are under /tests/contract/fixtures/layer3_adversarial)