Streamlit Cache Collision Vulnerability via Deterministic Sampling and PIL P-mode (CVE-2024-44728)

Vulnerability Overview This vulnerability involves two main issues within the Streamlit library: 1. Deterministic sampling seed leading to hash collisions: When a data object exceeds the sampling threshold, the code generates a smaller subset for hashing. Since the sampling seed is hardcoded, an attacker can pre-compute and inject malicious data, leading to cache poisoning. 2. Palette bytes of PIL P-mode images are ignored: For PIL P-mode images, the method returns only the palette indices of the pixels, not the actual color values. This causes visually distinct images to potentially produce identical hash values. Impact Scope Cache Poisoning: An attacker can pre-compute collision payloads by sending specific requests, thereby poisoning the cache so that subsequent users receive the attacker's data instead of the legitimate results. Data Integrity Issues: Two visually distinct images (e.g., a legitimate chart vs. a tampered image) may share the same cache key, resulting in incorrect cached outputs. Remediation 1. Sampling Seed: Derive the seed from the data itself (e.g., by hashing the first row or the first 64 elements) to ensure that the sampling indices depend on the content, preventing predictable collisions. 2. PIL P-mode: Call the method to retrieve the palette bytes before hashing, and include them in the hash computation. POC Code Bug 1 — Deterministic Sampling Seed Leading to Hash Collisions NumPy Example: Pandas Example: Bug 2 — Palette Bytes of PIL P-mode Images Are Ignored Environment Streamlit Version: All versions that include large object sampling (introduced via sampling optimization) Python Version: 3.10+ Operating System: Linux / macOS / Windows (Platform-independent) Conclusion The vulnerability has been confirmed, and the remediation patches have been submitted. The fix involves deriving the sampling seed from the data and including the palette bytes of PIL P-mode images before hash computation.

Goal Reached Thanks to every supporter — we hit 100%!

Streamlit Cache Collision Vulnerability via Deterministic Sampling and PIL P-mode (CVE-2024-44728)

More from this source