Security context
What an agent needs to avoid regressing past fixes and find the next vuln in this repo.
ggml-org/ggml
master @ 707321c
52
Fixes
16
CVEs
CRITICAL
Peak severity
40.0%
Coverage
Highlights
Memory Corruption: 17 prior fixes. Scrutinize any change in this area.
src/ggml.c: most-fixed (6 issues). Treat as high-risk during review.
36 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Remote Code Execution: Unvalidated CPU data pointers and unverified destination buffers during RPC tensor deserialization and copying allow unauthenticated attackers to write arbitrary process memory, leading to remote code execution. This is a highly critical threat vector in networked configurations.
Memory Corruption: Insecure GGUF parsing and tensor metadata initialization allow malformed or malicious model files to trigger integer overflows and heap-based buffer overflows, compromising memory safety prior to runtime computation.
Memory Corruption: Integer overflows in grid allocations, offset calculations, and missing boundary enforcement in parallel CUDA execution kernels can lead to out-of-bounds writes and corrupted execution states during matrix multiplications and reductions.