Unbounded memory allocation in highlight_lines range expansion in mdex
Get instant alerts when new vulnerabilities affect mdex - before attackers find them.
Memory Allocation with Excessive Size Value vulnerability in leandrocp mdex allows an unauthenticated attacker to cause a denial of service through unbounded memory allocation. comrak_nif::lumis_adapter::LumisAdapter::parse_highlight_lines in native/comrak_nif/src/lumis_adapter.rs eagerly expands a user-controlled inclusive line range from a fenced code block's highlight_lines decorator into a Vec<usize>, pushing one element per integer in the range with no upper bound on the range size. An attacker who can supply Markdown that an application renders with MDEx.to_html/2 (for example a comment, chat message, or wiki page) can embed a code block whose info string is rust highlight_lines="1-100000000", forcing the native adapter to allocate roughly 8 bytes per line in the range. A payload that differs by only a few bytes can therefore allocate hundreds of megabytes, and a sufficiently large range (for example 1-2000000000) exhausts host memory and aborts the BEAM, denying service to every user of the rendering process. The per-line write loop additionally tests membership with a linear scan over the same vector, degrading rendering to a quadratic cost even for ranges that do not immediately exhaust memory. The vulnerable native code originally shipped inside mdex (in native/comrak_nif/src/lumis_adapter.rs) and was later extracted into the separate mdex_native package (native/mdex_native_nif/src/lumis_adapter.rs), where it remains unpatched. This issue affects mdex from 0.11.0 before 0.12.3, and mdex_native from 0.1.0 before 0.2.3.
Our own answer: ranges merged, affected separated from platform-only, with confidence and review flags.
No remediation guidance has been written for this CVE yet. What the record does show:
SSVC signals action urgency (what to patch first), separate from CVSS which measures severity.
Get alerted
Track the products you run and VulnTracker tells you when a new CVE affects them — with what is known about fixing it, and nothing about the rest.