OpenMultimodalLab v1.1.2 is the current maintained release; v1.0.0 is the immutable research/evidence baseline. Security fixes land on main; backports and a patch release are decided case by case from impact and compatibility. No long-term-support branch is promised.

Before installing or updating the package in each environment, upgrade pip:

python -m pip install --upgrade "pip>=26.2"
python -m pip --version

pip versions below 26.2 are affected by CVE-2026-13346, which requires a malicious package index. Use a trusted index for the bootstrap as well as subsequent installations. The supported installation guides and CI apply this minimum separately to each environment, including newly created wheel-test venvs.

The Windows model constraints record the packages present during the frozen v1.0.0 license audit. The old pip entry is historical evidence, not a current installer requirement. Current installation commands do not consume this file. Preserve it together with its bound snapshot and report; see the maintenance policy. Do not use the historical file as a pip install -r installation list. A -c constraint does not itself install packages, but its old pip pin conflicts with the current pip minimum if pip is also requested. Record new environment audits under new names rather than overwriting the released evidence.

Do not open a public issue containing an exploit, credential, private media, personal path, or other sensitive evidence.

Use the repository Security tab and choose Report a vulnerability when private vulnerability reporting is available. If that option is unavailable, open a minimal public issue asking the maintainer to establish a private contact channel. Do not include vulnerability details in that issue.

Include privately:

• affected commit and platform;
• impact and realistic threat model;
• minimal reproduction with synthetic data;
• whether local files, model caches, credentials, or remote services are involved;
• suggested mitigation, if known.

No response-time or embargo SLA is promised. Please allow the maintainer to confirm impact and coordinate a fix before public disclosure.

• Core benchmark execution is local-first and does not upload media.
• No remote API backend is currently enabled by default.
• Model downloads use the repositories and immutable revisions documented by each backend.
• User outputs and model caches are excluded from Git.
• Dataset JSONL is limited to 16 MiB with 1 MiB per line; report JSONL is limited to 256 MiB with 4 MiB per record; resume manifests are limited to 8 MiB.
• Images are limited to 32 MiB and 40 million decoded pixels. Videos are limited to 256 MiB, 60 seconds, 3,600 source frames, and 3840×2160 pixels per frame before the fixed eight-frame sample is decoded.
• Durable run records keep relative media references and reduce absolute media paths to basenames. Windows drive, UNC, and POSIX absolute paths are redacted from persisted error text.
• The repository audit detects common credentials and personal paths, but it is not a complete secret-scanning or malware-analysis system.
• Images, documents, and videos remain untrusted native-parser inputs. The limits reduce resource-exhaustion exposure but do not sandbox Pillow, PyAV, FFmpeg, PyTorch, or Transformers. Use current dependency builds and process untrusted files in a disposable environment.

For non-sensitive defects, use the Bug report issue form and replace local paths, usernames, tokens, and private media with synthetic placeholders.