ChatGPT, Make a Secure Malloc for me
summary
Claims a workflow to synthesize production-grade allocators: map high-level allocation and security properties onto processor primitives (eg. CHERI capabilities), assemble a corpus of code fragments, and train a custom LLM with reinforcement learning plus model checking to produce correct, optimized mallocs. Motivates this by noting legacy allocators are huge (jemalloc ~29,000 lines; recent Microsoft allocator ~20,000) and that ~70% of security bugs stem from dynamic memory. Funding is secured and the project runs for three years to build and evaluate generated allocators.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:59Memory Allocation
- 1:53Secure Memory Management
- 2:37Dynamic Memory
- 3:05Cherry pointers
- 5:14Legacy allocator
- 6:07Handcraft allocators
- 6:39Synthesis
- 7:36Outro
related talks
37mHeap Layout Optimisation for Exploitation
19mStarMalloc: Verifying a Modern, Hardened Memory Allocator
22mGuarder: A Tunable Secure Allocator
14mScalable Memory Protection in the PENGLAI Enclave
22mHeapHopper: Bringing Bounded Model Checking to Heap Implementation Security
20mProtected DMAbufs and its dynamic memory assignment woes
44mHands Off and Putting SLAB/SLUB Feng Shui in a Blackbox
35mDeep-Kernel Treasure Hunt: Finding exploitable structures in the Linux kernel