Hello maintainers,
I am reporting six security findings that remain in the latest upstream main commit 06e47c70b76f419196678367dae62acca4c94076 (September 29, 2026). I compared the affected source files with the previously audited commit 33453cf123927f41c00e1935450e504d8a630763; all seven affected files are unchanged, so the original PoCs remain applicable.
The findings are:
-
Vision image SSRF and excessive buffering: src/utils/image.ts fetches arbitrary HTTP(S) image URLs with the CLI's network privileges and buffers the complete response before applying the nominal size limit. A caller who can influence the image URL may reach local/private services and cause excessive memory use.
-
Speech subtitle SSRF: src/commands/speech/synthesize.ts fetches the API-provided subtitle_file URL directly. A malicious or compromised response can make the CLI contact local or private network services.
-
Media download SSRF: src/files/download.ts accepts API-provided download URLs without rejecting local/private destinations, allowing an attacker-influenced response to make the CLI fetch and save unintended data.
-
Config temporary-file symlink overwrite: src/config/loader.ts uses a predictable config.json.tmp path. A local attacker able to plant a symlink can redirect credential/config data into another writable file.
-
Self-update temporary-file symlink overwrite: src/update/self-update.ts uses a predictable /tmp/mmx-update-<timestamp> path and normal write semantics. A local attacker can redirect update contents into another writable file.
-
Authenticated query-parameter injection: src/client/endpoints.ts interpolates task/file identifiers into URLs without encoding. A crafted identifier can add parameters to a credentialed request and change how the server interprets it.
The first three findings are network-request issues. The next two require local access to manipulate temporary files. The sixth requires an attacker-influenced or compromised identifier source. The local PoCs reproduced the behavior against clean upstream code using only loopback mock servers and temporary files. No production internal service, metadata endpoint, paid generation, upload, destructive request, API key, or bearer token is included here.
Proposed fixes and focused regression tests are available in a private local worktree. GitHub private vulnerability reporting is disabled for this repository, and issue #265 requested a private channel but has received no response. Please acknowledge this report, route it to the CLI/security maintainers, and provide a private channel for the complete PoCs and remediation discussion. Please also confirm whether MiniMax-AI/cli is in scope for a security reward or bounty.
I will keep the detailed exploit material private while waiting for your instructions.
Reporter: Jonathan Shi (x4evexnol)
Hello maintainers,
I am reporting six security findings that remain in the latest upstream
maincommit06e47c70b76f419196678367dae62acca4c94076(September 29, 2026). I compared the affected source files with the previously audited commit33453cf123927f41c00e1935450e504d8a630763; all seven affected files are unchanged, so the original PoCs remain applicable.The findings are:
Vision image SSRF and excessive buffering:
src/utils/image.tsfetches arbitrary HTTP(S) image URLs with the CLI's network privileges and buffers the complete response before applying the nominal size limit. A caller who can influence the image URL may reach local/private services and cause excessive memory use.Speech subtitle SSRF:
src/commands/speech/synthesize.tsfetches the API-providedsubtitle_fileURL directly. A malicious or compromised response can make the CLI contact local or private network services.Media download SSRF:
src/files/download.tsaccepts API-provided download URLs without rejecting local/private destinations, allowing an attacker-influenced response to make the CLI fetch and save unintended data.Config temporary-file symlink overwrite:
src/config/loader.tsuses a predictableconfig.json.tmppath. A local attacker able to plant a symlink can redirect credential/config data into another writable file.Self-update temporary-file symlink overwrite:
src/update/self-update.tsuses a predictable/tmp/mmx-update-<timestamp>path and normal write semantics. A local attacker can redirect update contents into another writable file.Authenticated query-parameter injection:
src/client/endpoints.tsinterpolates task/file identifiers into URLs without encoding. A crafted identifier can add parameters to a credentialed request and change how the server interprets it.The first three findings are network-request issues. The next two require local access to manipulate temporary files. The sixth requires an attacker-influenced or compromised identifier source. The local PoCs reproduced the behavior against clean upstream code using only loopback mock servers and temporary files. No production internal service, metadata endpoint, paid generation, upload, destructive request, API key, or bearer token is included here.
Proposed fixes and focused regression tests are available in a private local worktree. GitHub private vulnerability reporting is disabled for this repository, and issue #265 requested a private channel but has received no response. Please acknowledge this report, route it to the CLI/security maintainers, and provide a private channel for the complete PoCs and remediation discussion. Please also confirm whether
MiniMax-AI/cliis in scope for a security reward or bounty.I will keep the detailed exploit material private while waiting for your instructions.
Reporter: Jonathan Shi (
x4evexnol)