CVE-2025-14569: 
Linux Debian vulnerability analysis and mitigation

CVE-2025-14569 is a use-after-free vulnerability in ggml-org whisper.cpp affecting versions up to and including 1.8.2. The flaw resides in the read_audio_data function within /whisper.cpp/examples/common-whisper.cpp, triggered when ma_decoder_init_file is used to process audio files via the whisper-cli tool. The vulnerability was reported by a researcher on November 7, 2025 via a GitHub issue, and was publicly disclosed on December 12, 2025 with no vendor patch issued at the time of disclosure. It carries a CVSS v3.1 base score of 5.3 (Medium) (Red Hat CVE, VulDB).

The root cause is a use-after-free (CWE-416) combined with improper restriction of memory buffer operations (CWE-119) in the read_audio_data function of common-whisper.cpp. When ma_decoder_init_file initializes a decoder object on the stack, a subsequent free() call attempts to release a pointer that was not heap-allocated, resulting in an invalid free and crash. AddressSanitizer confirms the memory access occurs at offset 456 within the stack-allocated decoder variable (line 47 of common-whisper.cpp). Exploitation requires local access and low privileges — an attacker must be able to supply a crafted audio file to the whisper-cli binary. A public proof-of-concept reproducer file is available in the researcher's GitHub repository (GitHub Issue #3501, PoC Repo).

Successful exploitation can result in low-level impacts to confidentiality, integrity, and availability of the affected process, as reflected in the CVSS scoring. The invalid free causes an immediate application crash (denial of service), and depending on heap layout and platform mitigations, could theoretically be leveraged for memory corruption leading to limited information disclosure or code execution. The scope is limited to the local system and the whisper-cli process context, with no known lateral movement potential (Red Hat CVE, VulDB).

A public proof-of-concept reproducer file has been published by the original reporter in a GitHub repository, making exploitation accessible to any local user who can supply a crafted audio file to the whisper-cli binary. The EPSS score is very low at 0.013% (0.000130), indicating limited probability of widespread exploitation in the near term. There is no evidence of in-the-wild exploitation, no known threat actor attribution, and the vulnerability is not listed in the CISA Known Exploited Vulnerabilities (KEV) catalog. The CVSS v4.0 exploit maturity is rated as Proof-of-Concept (VulDB, ENISA EUVD).

1. Obtain local access: Gain a local user account on a system running a vulnerable version of whisper.cpp (≤1.8.2) with the whisper-cli binary installed.
2. Prepare a crafted audio file: Download or craft the public proof-of-concept reproducer file (available at the researcher's GitHub repository) designed to trigger the invalid free in ma_decoder_init_file.
3. Execute whisper-cli with the malicious file: Run the command ./install/bin/whisper-cli -m <model_file> -f ./repro, supplying the crafted audio file as input.
4. Trigger the vulnerability: The read_audio_data function calls ma_decoder_init_file, which initializes a decoder on the stack; a subsequent free() on the stack-allocated pointer triggers the invalid free, causing a crash or potential memory corruption.
5. Observe impact: The process aborts with an AddressSanitizer error or a segmentation fault, achieving denial of service; further exploitation for code execution would require additional heap-shaping techniques depending on platform mitigations (GitHub Issue #3501, PoC Repo).

• Process: Unexpected crash or abort of the whisper-cli process; AddressSanitizer output containing attempting free on address which was not malloc()-ed referencing ma_decoder_init_file or read_audio_data in common-whisper.cpp.
• Logs: System logs (e.g., /var/log/syslog) showing segmentation fault or abort signals from the whisper-cli process; core dump files generated in the working directory.
• File System: Presence of unusual or unexpected audio files (e.g., zero-byte or malformed files) supplied as input to whisper-cli; core dump files (core.*) in the application working directory.
• Network: No network-based IOCs expected given the local-only attack vector (GitHub Issue #3501).

As of the time of disclosure, the ggml-org project had not issued a patch or official response despite early notification via the GitHub issue tracker. Users should monitor the whisper.cpp GitHub repository for a fix and upgrade as soon as a patched version is released. As a workaround, restrict local access to the whisper-cli binary to trusted users only, and avoid processing untrusted or externally supplied audio files. Compiling with AddressSanitizer or other memory safety tools in development/testing environments can help detect exploitation attempts (GitHub Issue #3501, Red Hat CVE).

Red Hat has tracked the vulnerability in their security advisory database, though no Red Hat products are directly affected. The vulnerability received automated coverage from threat intelligence aggregators and CVE tracking services shortly after disclosure. No notable researcher commentary, vendor statements beyond tracking, or significant media coverage has been identified beyond the original GitHub issue report (Red Hat CVE).