aggressive safetysettings (for example block_low_and_above) on red-team or content-moderation workloads can produce frequent safety finishes. exclude those projects or models, or require concurrent high/medium ratings.

Techniques

Sample rules

GCP Vertex AI Response Blocked by Safety Filters

source: elastic
technicques:

Description

Detects Vertex AI GenerateContent responses whose candidate finish_reason is SAFETY. That value is emitted when Gemini safety filters stop generation. Unlike textual refusals (finish_reason STOP), SAFETY means the provider blocked or truncated the completion.

Detection logic

from logs-gcp_vertexai.prompt_response_logs-* metadata _id, _version, _index
| where
    data_stream.dataset == "gcp_vertexai.prompt_response_logs" and
    MV_CONTAINS(gcp.vertexai.prompt_response_logs.full_response.candidates.finish_reason, "SAFETY")
| keep
    _id,
    _version,
    _index,
    @timestamp,
    cloud.project.id,
    gcp.vertexai.prompt_response_logs.model,
    gcp.vertexai.prompt_response_logs.api_method,
    gcp.vertexai.prompt_response_logs.full_request.contents.parts.text,
    gcp.vertexai.prompt_response_logs.full_request.safety_settings.category,
    gcp.vertexai.prompt_response_logs.full_request.safety_settings.threshold,
    gcp.vertexai.prompt_response_logs.full_response.candidates.finish_reason,
    gcp.vertexai.prompt_response_logs.full_response.candidates.safety_ratings.category,
    gcp.vertexai.prompt_response_logs.full_response.candidates.safety_ratings.probability,
    gcp.vertexai.prompt_response_logs.full_response.candidates.safety_ratings.severity