Uncontrolled Recursion in llama.cpp

Executive Summary

llama.cpp versions up to b5693 have a flaw in common/json-schema-to-grammar.cpp where deeply nested JSON schemas cause uncontrolled recursion. This crashes the server by exhausting the recursion stack, leading to a denial of service.

Useful 0 Not Useful 0

Detection Guidance

To detect this vulnerability, monitor for crashes in llama.cpp processes handling POST /completions requests. Check server logs for stack overflow errors or segmentation faults during JSON schema processing. Use tools like strace or gdb to trace recursive calls in SchemaConverter::visit and _generate_union_rule functions. Test with a deeply nested JSON schema (depth 7000+) to trigger the issue.

Useful 0 Not Useful 0

Impact Analysis

An attacker could send a specially crafted JSON schema to the POST /completions endpoint, causing the server to crash. This disrupts service availability for all users relying on that endpoint.

Useful 0 Not Useful 0

Mitigation Strategies

Immediately disable the POST /completions endpoint if not required. Implement input validation to reject excessively nested JSON schemas. Apply network-level rate limiting to prevent abuse. Monitor for unusual traffic patterns targeting the vulnerable endpoint. Consider upgrading to a patched version if available.

Useful 0 Not Useful 0