# Utah security workbench Open /utah/security/workbench with your existing Utah account. Only previously verified targets can be assessed. API: /utah/api/security/workbench (or /api/security/workbench on the account service). Use your agent execution bearer key; model-only keys cannot read or execute security work. GET the base path to list missions. POST {request_id,input:{title,host}} to create one. GET /{mission_id} for the graph, findings, coverage, limits and history. GET /{mission_id}/export for the full report and canonical JSON SHA-256. GET /{mission_id}/evidence/{evidence_id} for retained, redacted evidence. GET /{mission_id}/search?q=... to search traces and evidence. POST /{mission_id}/{operation} with {request_id,input}. Use a unique stable request_id per operation; reuse it only for the same input. Operations: assets, step, plan, run, observe, probe, adopt, capture_start, capture_ingest, capture_stop, finding, coverage, retest, retest_run, resume, archive, purge. step input: {title,tool,url,hypothesis,control,requires:[],files:[{path,content}]}. run/observe input: {id:step_id}. plan reuses Utah's deterministic adaptive planner; it does not launch a job. probe input: {url,method:"GET"|"HEAD"}. No session credentials are forwarded. Redirects are not followed. capture_start input: {minutes:15}; ingest: {capture_id,entries:[{url,method,status,request,response,headers,complete}]}. finding input: {title,severity,asset_id,risk,description,evidence:[{id,role:"proof"|"control"|"baseline",quote}],reproducible_http:false}. retest input: {finding_id,revision,preconditions}; retest_run input: {id:retest_id}. A target cooldown may require a later retry. coverage input: {asset_id,risk,outcome:"blocked"|"inconclusive"|"needs_follow_up"|"not_applicable",reason}. HTTP 429 is target pacing. Keep the saved step/retest and try again after cooldown; do not invent a replacement launch. Fixed requires complete matching server-observed trigger and normal-control responses. An authentication failure, truncation or absent scanner match is inconclusive. Authenticated HAR imports remain labeled customer supplied; they cannot promote themselves to server observations. Path-limited scopes use read-only probes or source tools; domain scanners cannot be confined to a path. Source tools require at most 60 KB of supplied files. Raw source is encrypted and is not returned in mission reports. Mission evidence is capped at 500 records / 5 MB. Raw unreferenced evidence expires after 1–90 days; finding/retest references retain their evidence until mission purge or account deletion. Report export is always available after the work budget expires. Owner acceptance is separate from all automated checks. MCP: utah_security_workbench, with operation, mission_id, request_id and input. It calls these same routes. Agent action: {action:"security",operation:"workbench",command:"read",mission_id:"mission_..."}. Terminal: download /utah/security/cli.py and run python3 cli.py --help. Set UTAH_API_KEY in your environment; all state and execution remain hosted.