Until v1.0, only the latest main revision receives security fixes.
Please use GitHub private vulnerability reporting when available. Do not disclose exploitable model-loading, artifact-path, deserialization, or checkpoint-overwrite issues publicly before maintainers have assessed them.
Model and dataset inputs are untrusted. Remote model code is disabled by default. ModelSurgeon must not overwrite source checkpoints by default, must use atomic output publication, and must constrain artifact paths to the configured run directory. Generated models can behave unexpectedly and require evaluation before use.
The v2.9 conversational security gate is a bounded, local, typed control-plane contract; it is not a general security certification. Consequential actions require scoped approval for the exact plan and material diff. Prompts, provider text/metadata, and tool output are untrusted and cannot override hard constraints, approvals, evidence, transactions, or source-model immutability. Unknown, contradictory, malformed, stale, and isolation-failure states fail closed, and negative or unresolved observations remain retained.
The repository does not claim operating-system containment for hostile in-process providers or handlers, live hosted-provider or credential security, complete future prompt-injection resistance, distributed campaign recovery, optimizer optimality, or universal deployment/model-family support. Do not put credentials in prompts, fixtures, evidence bundles, diagnostics, or chat history. See the v2.9 security boundary and residual-risk record. The v3.0 product release preserves these limits and audits them through the v3.0 release manifest; it is not a general security certification.
The core package does not depend on diskcache and does not import it. The
optional local GGUF adapter requires a separately managed llama-cpp-python
runtime; current releases of that runtime pull unmaintained diskcache, which
uses pickle serialization by default and has no patched release. Do not expose
that runtime's cache directory to untrusted writers, and do not treat the
optional local provider as production-safe until its dependency graph is
replaced or patched.