Run Gitleaks against every Git ref before publishing a source or migration archive:
gitleaks detect --source . --log-opts=--all --redact --no-bannerThe checked-in .gitleaks.toml extends the default detector set and narrowly
allowlists two reproducibility identifiers: the SHA-256 identity of the
committed tokenizer and its literal byte-tokenization protocol name. Neither is
a credential. New hashes and protocol identifiers remain subject to the default
detectors.
flaxchat.execution.execute_code is a best-effort reliability guard for code
that an operator has already reviewed and trusts. It is not a security
sandbox. The default is fail-closed: calls must pass trusted=True.
The guarded child process has a wall-clock timeout, bounded captured output, best-effort Linux memory/process/file limits, an empty environment, and a temporary working directory. These controls protect the parent from common accidents; they do not defend against a malicious Python program, native code, kernel exploits, filesystem traversal, or all network/import bypasses.
Production systems must run untrusted generated code in a separately reviewed
container, microVM, or remote worker. Flaxchat ships an opt-in Docker/Podman
adapter, execute_code_isolated, that requires a digest-pinned image and uses
no network, no inherited credentials, no host mounts, a read-only root, an
unprivileged user, dropped capabilities, no-new-privileges, isolated Python,
CPU/memory/process/output/time quotas, and bounded disposable tmpfs storage.
The container runtime and kernel remain part of the trusted computing base.
HumanEval is fail-closed unless FLAXCHAT_EXECUTION_IMAGE names a reviewed
image using repository@sha256:<64 lowercase hex characters>. Set
FLAXCHAT_EXECUTION_RUNTIME=podman to use Podman; Docker is the default. Never
replace the digest with a mutable tag. Deployments needing a stronger boundary
should substitute a reviewed microVM or remote-worker backend.