CVE-2026-43632
Use-after-free in llama-server tokenization endpoints enabling crash or RCE
Is CVE-2026-43632 being exploited?
Not confirmed. CVE-2026-43632 does not appear in CISA's Known Exploited Vulnerabilities catalog, which records only exploitation that has been observed and reported publicly. That is evidence of absence of a report, not evidence the flaw is unattacked. EPSS currently estimates a 0.33% probability of exploitation in the next 30 days.
How severe is CVE-2026-43632?
CVE-2026-43632 is rated Critical with a CVSS score of 9.2. Severity describes how bad exploitation would be, not how likely it is: pair it with exploitation evidence before deciding what to patch first.
Is there a patch for CVE-2026-43632?
Not known from our data. No patch reference has been recorded for CVE-2026-43632, which is not the same as no patch existing — a fix may have shipped without a tagged reference, or after this record was written. The vendor advisory is the only authority on whether a fix exists, and mitigations may be available regardless.
What does CVE-2026-43632 affect?
CVE-2026-43632 affects llama.cpp builds b7492–b9060, llama-server, Endpoints: /tokenize, /detokenize, /infill, /apply-template, /rerank, /anthropic/count_tokens. Confirm the exact affected versions against the vendor advisory before deciding you are exposed.
What should I do about CVE-2026-43632?
Isolate/restrict endpoints, disable --sleep-idle-seconds, apply vendor patch when released
Exploitation status reflects CISA's KEV catalog as we last synced it. Check the catalog directly.
Remote crash or potential remote code execution via TOCTOU use-after-free in HTTP tokenization endpoints when --sleep-idle-seconds is enabled.
Immediate action required
- affected>= b7492 and <= b9060
As published in the CVE Program record. A version outside these ranges is not a statement that it is unaffected — vendors sometimes understate a range, and distribution-backported builds carry upstream numbers that do not reflect what was patched into them.