$ the-wire · showcase
DockerDeployer env stops overriding governance settings, Guardrails validates tool-rail models at load
By RepoJournal · Filed · About Agent Security · Composed from the cited sources · methodology
Two fixes for operators of the agent governance stack: the Agent Governance Toolkit runtime now refuses a container deploy whose env would silently disable the kill switch, and Guardrails catches a missing judge model when config loads instead of on the rail's first run.
fix(agent-runtime): stop DockerDeployer env from overriding governance settings (#4409) microsoft/agent-governance-toolkit
DockerDeployer used to put the AGT_* variables derived from GovernanceConfig on the docker run line before your own env entries, and since Docker keeps the last -e for a name, env={"AGT_KILL_SWITCH": "false"} started the container with the kill switch off even when config.kill_switch_enabled was True; the same held for the policy path, trust level, audit flag and limits. deploy now raises Va[se...
feat(config): Check model references in tool rails NVIDIA-NeMo/Guardrails
A tool rail whose manifest binds a model, such as a judge named by $model= on tool safety check, used to go unchecked against your configured models, so a missing one only surfaced when IORails built the engine or on the rail's first run. The new RailsConfig validator check_model_exists_for_tool_rails covers every flow in rails.tool_call and rails.tool_result, global and per_tool, so a typo in ...
ci: add generate_workflows.py --sync-registry to adopt Dependabot action bumps microsoft/agent-governance-toolkit
generate_workflows.py gains a --sync-registry mode that reads a bumped pin back out of the generated workflow files and updates .github/ci/actions.toml, the source of truth Dependabot cannot edit; previously an actions major left policy-engine-ci.yml drifting from the registry and --check failing until a maintainer hand-synced it. Relevant only if you maintain the pipeline, but it keeps the reg...
