llama.cpp: Out-of-bounds Read
CVE-2026-43628 is a high-severity out-of-bounds read vulnerability in llama.cpp from Ggml-org, with a CVSS 3.1 base score of 7.8 out of 10. It can be exploited with local access to the system by an unauthenticated attacker, but only if a user takes part, for example by opening a link or file. Affected llama.cpp versions: b3978 through b9058. The record does not name a fixed version, so check the vendor advisory. There is no confirmed exploitation so far. EPSS estimates a 0.23% chance of exploitation within 30 days, which is low compared with other CVEs.
High severity with no evidence of exploitation so far.
What to do
- Check which version of llama.cpp you run against the affected versions below, and install the vendor's latest release.
- Because the attack needs a user to take part, warn users about unexpected links and files related to this product until it is patched.
Affected products and versions
Version data comes from the organisation that assigned the CVE. Confirm against the vendor advisory before deciding you are unaffected.
How the severity score breaks down
- Attack vector
- LocalThe attacker needs local access or must trick a user into running something.
- Attack complexity
- LowNo special conditions are needed; the attack works reliably.
- Privileges required
- NoneNo account or login is needed.
- User interaction
- RequiredA user must take an action such as opening a file or link.
- Scope
- UnchangedImpact stays within the vulnerable component.
- Confidentiality
- HighAll data handled by the component can be read.
- Integrity
- HighData or code can be modified at will.
- Availability
- HighThe component can be taken down completely.
Vector: CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H. Scored by CNA (vulncheck.com).
| Scoring system | Score | Severity | Scored by |
|---|---|---|---|
| CVSS 4.0 | 8.5 | High | CNA (vulncheck.com) |
| CVSS 3.1 | 7.8 | High | CNA (vulncheck.com) |
Official description
llama.cpp builds b3978 through b9058 contain an integer underflow and out-of-bounds read vulnerability in the DRY sampler that allows unauthenticated attackers to trigger a heap buffer underflow by sending a crafted HTTP request with dry_allowed_length set to INT32_MIN to the /v1/completions or /v1/chat/completions endpoints. Attackers can exploit this vulnerability to crash the server with SIGSEGV causing denial of service for all connected users, or corrupt token sampling probabilities by reading garbage values from memory before the allocated buffer.
Source: the CVE record as published by its numbering authority, via NVD and CVE.org.
The underlying weakness: Out-of-bounds Read
The product reads data past the end, or before the beginning, of the intended buffer.
How developers prevent this class of flaw
- Use a language that provides appropriate memory abstractions.
Timeline
- CVE record published.
- CISA published its SSVC assessment (exploitation: none, automatable: no, technical impact: total).
- Record last changed materially (score, affected products, status or references).
References
Patches and release notes
- github.com/Vladimir-tokarev-cyera/llama-cpp-security-patchesThird Party Advisory
- vulncheck.com/advisories/llama-cpp-b3978-b9058-integer-underflow-via-dry-samplerThird Party Advisory
Other vulnerabilities in llama.cpp
- 7.5 High severity
- 8.1 High severity
- 7.8 High severity
- 8.2 High severity
- 8.1 High severity
- 7.3 High severity
Recent Out-of-bounds Read vulnerabilities
- 7.1 High severity
- 8.1 High severity
- 7.1 High severity
- 7.5 High severity
- 7.8 High severity
- 9.8 Critical severity
Spotted an error? Tell us.