$ the-wire · showcase
India policy pack denies everything, and rollouts silently drop analysis
By RepoJournal · Filed · About Agent Security · Composed from the cited sources · methodology
Operators of Microsoft's agent-governance-toolkit have two live problems: the India regulatory example policies return deny on every evaluation, and the AgentRollout CRD prunes the analysis criteria that decide whether a rollout step passes.
fix(examples): india cert-in and sebi policies deny every evaluation microsoft/agent-governance-toolkit
The certin-2022-directions and sebi-governance rules in examples/policies/india-regulatory tested a parameter with a negated comparison instead of binding it, and a negated comparison against a missing path is true. Anything you gate on those policies, benign content included, is being denied right now.
fix(examples): india cert-in and sebi policies deny every evaluation (#4480) microsoft/agent-governance-toolkit
The fix binds the parameter properly, but the merge note warns these bundles are bound to runtime intervention points, where an absent parameter is not an unusual input. Re-pull the example pack and re-run your decision probes before trusting either policy in a live hook.
fix(agent-sre): keep a rollout step's analysis criteria in the AgentRollout CRD (#4476) microsoft/agent-governance-toolkit
The AgentRollout CRD schema in both generate_crd_manifest() and the Helm chart never declared steps[].analysis, so the API server pruned it from every applied rollout and the Reconciler built steps with no analysis criteria. Analysis is now declared as a list of metric/threshold/comparator, with a test that the Helm schema matches the generated one.
docs: Relay client plugin NVIDIA-NeMo/Guardrails
The docs for the Rust nvidia.nemo_guardrails plugin for NeMo Relay now cover how to build, install and configure it, where before they described only the /v1/hooks/nemo-relay endpoint it calls. Your tool policy and rails decide whether a Relay-managed call runs, and now you can wire that up without reading the plugin source.
fix(server): support LiteLLM translated response metadata NVIDIA-NeMo/Guardrails
LiteLLM-translated Chat responses and streams keep reviewed metadata and encrypted reasoning items instead of tripping the closed guard contracts, while unsupported content is still rejected. If you front models through LiteLLM, check your gateway recordings against the new declarations before relying on prior verdicts.
Action items
- → Update the certin-2022-directions and sebi-governance policies from microsoft/agent-governance-toolkit and re-run a b... microsoft/agent-governance-toolkit [plan]
- → Re-apply AgentRollout manifests after regenerating the CRD or upgrading the Helm chart, since steps[].analysis was pr... microsoft/agent-governance-toolkit [plan]
