Current version: 0.1.0. The canonical value lives in src/sdk/ts/package.json; scripts/check_version.js fails CI if this table drifts from it.

TakyonDB is pre-alpha. Do not expose the daemon or shared-memory segment to untrusted networks or processes, and do not rely on it for data you cannot lose yet: it has no replication, and a single segment is a single disk's worth of durability.

• The Zig core uses mmap / CreateFileMapping shared memory and raw pointer arithmetic. All C-ABI entry points validate offset/size/ key_len, and scripts-driven fuzzing covers the gated entrypoints and the pure kernels, but the covered surface is not the whole surface.
• The Node-API addon (binding.cc) runs in-process. Treat malformed ArrayBuffer offsets as untrusted input.
• The WAL uses CRC32 to detect torn writes, not cryptographic integrity. It does not protect against malicious tampering: an attacker who can write to the data directory can write a log that passes its own CRC.
• Every client maps the arena read-write, including readers. There is no read-only mapping in practice, so a client process that is compromised can modify any record any other client can see.
• The admin endpoint binds 127.0.0.1 and speaks line-based ASCII with no authentication. Anything that can reach that port can read keys through SCAN and RANGE, and can trigger a checkpoint.
• The relational catalog and rows live in the same arena with the same bounds checks as the key-value path; there is no additional isolation between models. Today the rows are in the SDK's own heap, so the relational surface is the JavaScript one, and its checks are its own.
• A where clause cannot reach the engine: predicates are evaluated in the SDK against objects it already holds, so the C ABI is not on that path.

Use GitHub's private vulnerability reporting: Security → Report a vulnerability on this repository. It opens a private thread with the maintainer, so the details never land in the public issue tracker.

This page used to say "open an issue with the security label". The label did not exist, which meant the policy was a dead end: the one thing a reporter should never do, put a vulnerability in public, was what the policy asked for. The label exists now, for issues that are about security without being exploits, and private reporting is the path for the rest.

Please include:

1. Affected version or commit
2. OS and architecture
3. Minimal reproducer, as Zig or TypeScript
4. Impact assessment, and whether the data at risk is reachable from an untrusted party

Reports are acknowledged within 72 hours. There is no bounty and no disclosure SLA beyond that; this is an early project with one maintainer, and pretending otherwise would be worse than saying it.