The tools
Both are read-only, so they run without an approval prompt even in manual approval mode. In Agent mode, like every tool, they run once a step of an approved plan is in progress.
query_attack_surface replaces the earlier per-topic lookups. Agent definitions and settings that still name an old tool keep working: the name now points to query_attack_surface.
Full details of one finding — description, proof of concept, remediation, evidence, timeline — still come from
investigate_vulnerability, and the dashboard score and trends from gather_statistics.How the agent learns the language
- The tool description carries the rules and the exact field names of every resource, generated from the same catalog the server validates against, so the model never has to guess a name.
- The
attack-surface-queryskill (see Skills) teaches the full language — operators, time, aggregates, paging, limits, the errors and how to fix them — with ready-made queries. The agent loads it before its first query of a conversation, and every built-in agent declares it. - Errors are written for the model: an unknown field lists the valid fields of that resource, a field that cannot be grouped lists the ones that can, and a query over budget says what to change. The agent fixes the whole call and sends it once more.
Permissions and scope
- The agent acts for the member running the conversation: each resource requires that member’s read permission (
asset.read,vulnerability.read,target.read). - A project chat locks every query to the project’s assets; a
projectIdthe model passes cannot leave it. - On the MCP server, which authenticates with the workspace API key and has no member, the tool reads with
asset.read,vulnerability.readandtarget.read.
Limits
The agent and MCP clients run under tighter limits than the API: 5 queries per call, 200 rows per query, 3 includes, 20 children per include, 20 conditions per filter, 3 seconds per statement, 64 KB per answer, and no large HTTP fields (body, headers). Report mode is not available to them. Everything else in the language is theirs too: nested filters, relation tests, conditional aggregates, having and filled time series. See Paging, limits and cost.
Using it from an MCP client
Connect the client as described in MCP Server, then let the assistant calldescribe_attack_surface and query_attack_surface. A call looks like:
{ "error": { "code", "message", "issues", "hints" } }, so the assistant can read them and retry.
