Please report suspected vulnerabilities privately rather than in a public issue. Use GitHub's Security advisory - Report a vulnerability form on this repository (Security tab - Report a vulnerability). Do not open a public pull request or issue for an unfixed vulnerability.

Please include: affected module and version (3.0.0-dev, git rev-parse HEAD), the platform, a minimal reproducer, and the impact (crash, memory corruption, out-of-bounds read/write, file-write, code execution, DoS).

Target response times:

Disclosure is coordinated with the reporter. A CVE is requested for issues that affect users of the released library; everything is credited to the reporter unless anonymity is requested.

This repository tracks the OpenCV 3.0 development line. Fixes land on the default branch (master); backports to other branches are not provided.

OpenCV parses attacker-controlled data in several places. When you review or patch code in these areas, treat every length, offset and count read from a file as hostile:

• modules/imgcodecs/src/grfmt_*.cpp - image header parsers (BMP, GIF, JPEG, PNG, TIFF, RAS, PXM, ...). Header fields drive allocations and loop bounds.
• modules/imgcodecs/src/bitstrm.cpp, loadsave.cpp - stream reads.
• modules/core/src/persistence.cpp - the XML/YAML reader behind cv::FileStorage. Never feed it a file from an untrusted source, and prefer FileStorage in read mode only.
• modules/videoio/src/cap_*.cpp - container/stream parsers for FFmpeg, GStreamer and image sequences.
• modules/core/src/*.cpp - Mat indexing, ROI/submatrix arithmetic, and the resize/remap geometry code; sizes come from the caller and are easy to overflow.

Rules the codebase follows, which new code must follow too:

1. Validate at the boundary. Public entry points check types, sizes, channel counts and non-negativity with CV_Assert/CV_Error and throw cv::Exception; they never trust a size that came from a file or a pointer arithmetic result.
2. Bounded copies. Use memcpy with a length that has been validated against both buffer sizes, or the cv::/std:: helpers that do. No unbounded strcpy/sprintf on data derived from an input file.
3. Checked arithmetic. Compute allocation sizes and pixel counts in a type that cannot silently wrap (e.g. size_t, or cv:: size helpers) before allocating.
4. No undefined behaviour. No uninitialised reads, no signed overflow, no out-of-bounds access, no reliance on NULL checks that do not happen.
5. Deterministic errors. Report failures through CV_Error with a specific Error::St* code rather than asserting or aborting, so callers can distinguish "bad input" from "out of memory".

• Every third-party dependency is either vendored under 3rdparty/ or pinned exactly in a requirements file (requirements-dev.txt, modules/python/requirements.txt); nothing floats on a version range.
• CI (.github/workflows/ci.yml) runs with permissions: contents: read, so a compromised action or dependency cannot push, release or read secrets from a job.
• CI installs the toolchain it needs explicitly and audits the pinned Python dependencies with pip-audit --strict on every pull request.
• Dockerfile builds from a pinned base image tag with a minimal package set and no credentials; .dockerignore keeps .git, local .env files, keys and build trees out of the build context and the image.
• Dependabot (.github/dependabot.yml) keeps the actions, container base and pinned Python tooling current.
• There are no secrets in this repository. .env.example documents the build knobs; the real .env is git-ignored, and any future credential must come from the CI secret store or the environment, never from a committed file.