BEGIN:VCALENDAR
VERSION:2.0
PRODID:Data::ICal 0.24
BEGIN:VEVENT
DESCRIPTION:   'Title: Breaking Local AI Runtimes: Exploiting llama.cpp and
  Ollama\n   Tags: DEF CON Official Talk | Demo 💻 | Exploit 🪲\n   Whe
 n: Friday\, Aug 7\, 10:30 - 11:30 PDT\n   Where: LVCCW Level 1 Hall 3 1007
  (Main Track 2) and DCTV-2 - [1]Map\n\n   Description:\n\n   Local LLM run
 times now sit inside phones\, desktops\, and internal\n   servers\, but th
 e layer underneath is still ordinary native code. We\n   analyzed llama.cp
 p and Ollama across three trust boundaries: JNI\, HTTP\n   lifecycle code\
 , and Go/C bindings.\n\n   First\, in the llama.cpp Android integration\, 
 Java can free a native\n   llama_context while native code is still using 
 it. We reclaim the\n   freed 648-byte object\, redirect a vtable call\, an
 d show code execution\n   in the embedding app. Second\, in llama.cpp serv
 er\, idle model teardown\n   can race active requests\, leaving a dangling
  pointer inside a freed\n   17\,816-byte model allocation. We show remote 
 cross-thread reclaim and\n   attacker-controlled native dereference\, then
  explain the remaining\n   steps to stable RCE. Third\, in Ollama\, malici
 ous GGUF metadata can\n   push unsafe lengths across the Go/C boundary dur
 ing quantization\,\n   causing C to read past a Go-backed buffer and retur
 n heap data to the\n   caller.\n\n   This is not a prompt-injection talk. 
 It is about exploiting local AI\n   runtimes as native software: one full 
 exploit\, one validated\n   server-side primitive\, one disclosure primiti
 ve\, and the audit\n   patterns that find more.\n\n     1. \n\n       Merg
 endahl\, Louloudis\, Vidas. "Cross-Language Attacks." NDSS\n       Symposi
 um 2022.\n\n     2. \n\n       Hussain. "Incubated Machine Learning Exploi
 ts." DEF CON 32\, 2024.\n\n     3. \n\n       Riancho\, Braverman\, Demetr
 io. "Breaking Out of The AI Cage." Black\n       Hat USA 2025.\n\n     4. 
 \n\n       llama.cpp project: https://github.com/ggml-org/llama.cpp\n\n   
   5. \n\n       Ollama project: https://github.com/ollama/ollama\n\n     6
 . \n\n       llama.cpp Android sample:\n       https://github.com/ggml-org
 /llama.cpp/tree/master/examples/llama.android\n\n   Speakers:Ofek Itach\,V
 ladimir "G1ND1L4" Tokarev\n\n   SpeakerBio:  Ofek Itach\, Cyera\n\n   Secu
 rity Research Team Lead at Cyera. Focus areas include cloud\n   infrastruc
 ture and AI-related platform security. Talks: Black Hat USA\n   2024\, DEF
  CON 32 (2024)\, RSA 2024\, Sector 2024\, INTENT 2024\, Black Hat\n   Euro
 pe 2024 Arsenal. Earlier work includes AWS internals and cloud\n   attack 
 surface mapping.\n\n   SpeakerBio:  Vladimir "G1ND1L4" Tokarev\, Cyera\n\n
    Vladimir Tokarev is a vulnerability researcher tech lead at Cyera\,\n  
  specializing in Cloud\, IoT/OT\, Windows\, Linux\, and AI vulnerability\n
    research and exploit. Talks: Black Hat USA 2024 and 2023\,\n   DEF CON 
 33 Recon Village 2025\, CodeBlue 2025\, RSA 2024.\n\n   '\n\n   1. #LVCCW_
 Level1_Hall3\n\n\n
DTEND:20260807T183000Z
DTSTART:20260807T173000Z
LOCATION:DEF CON Talks - LVCCW Level 1 Hall 3 1007 (Main Track 2) and DCTV-
 2
SUMMARY:Breaking Local AI Runtimes: Exploiting llama.cpp and Ollama
END:VEVENT
END:VCALENDAR