Platform

Security review workflow

POST /v1/workflows/security-review
GET  /v1/workflows/templates

One call runs a full review:

  1. Intake. One pass summarizes the material and lists what is in scope.
  2. Reviewers. Up to 12 reviewer personas run in parallel. Each returns findings with a location, description, fix and confidence.
  3. De-duplication. Findings from different reviewers are merged when their locations agree and their titles overlap. This step is deterministic and uses no model.
  4. Aggregation. A lead-reviewer pass decides, for each merged finding, whether to keep it and what its final severity is.

Every pass is an ordinary metered request. It can name its own model or an auto alias, and the response lists the cost of each pass. The call is synchronous and can take up to 5 minutes.

Templates

curl https://api.utmost.site/v1/workflows/templates

No key is needed. Each template comes back as YAML you can copy and edit.

Template Reviewers Models
code-review injection, authz, secrets-crypto, logic Reviewers on auto, aggregation on auto-best
threat-model stride, attacker, insider-supply-chain stride and attacker on auto-best, the rest on auto, aggregation on auto-best
dependency-audit known-vulns, supply-chain, license All on auto

Request

curl https://api.utmost.site/v1/workflows/security-review \
  -H "Authorization: Bearer $UTMOST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "template": "code-review",
    "input": {
      "content": "app.get(\"/user\", (req, res) => db.query(\"select * from users where id=\" + req.query.id))",
      "filename": "api/users.js",
      "language": "javascript",
      "context": "Public Express API behind a load balancer."
    }
  }'
Field Notes
template code-review, threat-model or dependency-audit
config A YAML string (up to 64,000 characters) or a JSON object. See the schema below.
input A string, or {content, filename?, language?, context?}. content is up to 600,000 characters, context up to 20,000.

Send template, config, or both. With both, the top-level keys in config replace the template's. The merge is shallow: a reviewers list in config replaces the template's whole list.

The request counts as one request against your rate limit; the passes inside it do not. Each pass checks your balance and spend caps, and goes through the hard-limit checks.

If you send x-utmost-variant, it applies to every pass that uses an auto alias.

Config schema

name: api-review              # optional, up to 64 characters
intake:
  model: auto                 # default: auto
  instructions: >-            # optional; replaces the default intake instructions
    Identify the HTTP handlers and the data each one touches.
reviewers:                    # required, 1 to 12
  - name: injection           # 1 to 64 characters
    model: auto               # default: auto
    persona: >-               # 1 to 8,000 characters
      You are an application security engineer. Find injection flaws.
      Trace untrusted input to sinks.
  - name: authz
    model: utmost/kimi-k3.1-pro
    persona: You review authentication and authorization.
aggregate:
  model: auto-best            # default: auto-best
  instructions: Be strict about false positives.   # optional
min_confidence: 0.5           # 0 to 1, default 0; drop findings below this
max_findings: 50              # 1 to 500, default 100

model takes any model id your key may call, or an auto alias. Pinning reviewers to abliterated models needs a Verified organization.

Response

{
  "id": "wfr_e7Kdbc4kflCHW3TxK0HF",
  "object": "workflow.run",
  "workflow": "security-review",
  "template": "code-review",
  "config_name": "code-review",
  "status": "completed",
  "duration_ms": 18234,
  "intake": { "summary": "An Express handler that reads a user by id.", "language": "javascript", "scope": ["SQL query construction"] },
  "findings": [
    {
      "id": "F1",
      "title": "SQL injection through req.query.id",
      "severity": "critical",
      "location": "api/users.js:1",
      "description": "The id is concatenated into the SQL string.",
      "recommendation": "Use a parameterized query.",
      "confidence": 0.96,
      "found_by": ["injection", "logic"],
      "duplicates_merged": 1,
      "aggregate_rationale": "Direct concatenation of a query parameter; exploitable."
    }
  ],
  "summary": { "critical": 1, "high": 0, "medium": 0, "low": 0, "info": 0 },
  "passes": [
    { "name": "intake", "kind": "intake", "model": "utmost/qwen3-30b-a3b", "route": "auto->small/standard", "request_id": "req_ZgMP...", "usage": { "input_tokens": 72, "cached_tokens": 0, "output_tokens": 40 }, "cost_usd": 0.000018 },
    { "name": "injection", "kind": "reviewer", "model": "utmost/deepseek-v4.1-flash", "route": "auto->medium/standard", "request_id": "req_ztCn...", "usage": { "input_tokens": 190, "cached_tokens": 0, "output_tokens": 210 }, "cost_usd": 0.000309, "findings": 1 },
    { "name": "aggregate", "kind": "aggregate", "model": "utmost/kimi-k3.1-pro", "route": "auto-best->large/standard", "request_id": "req_pVlX...", "usage": { "input_tokens": 420, "cached_tokens": 0, "output_tokens": 90 }, "cost_usd": 0.00174, "findings": 1 }
  ],
  "cost": {
    "total_usd": 0.002067,
    "per_pass": [
      { "name": "intake", "model": "utmost/qwen3-30b-a3b", "cost_usd": 0.000018 },
      { "name": "injection", "model": "utmost/deepseek-v4.1-flash", "cost_usd": 0.000309 },
      { "name": "aggregate", "model": "utmost/kimi-k3.1-pro", "cost_usd": 0.00174 }
    ]
  }
}

(The passes list above is shortened; a code-review run has one entry per reviewer.)

Findings

Field Meaning
severity critical, high, medium, low or info. The aggregation pass can change it.
confidence Combined across reviewers that found it: 1 − ∏(1 − cᵢ), using each reviewer's highest confidence. Agreement raises it.
found_by Reviewer names
duplicates_merged How many other findings were merged into this one
aggregate_rationale The lead reviewer's reason, or null if aggregation did not run or did not cover it

Findings the aggregation pass drops are removed. Then findings below min_confidence are removed, the rest are sorted by severity and confidence, and the list is cut to max_findings. summary counts the final findings by severity.

Passes and partial failure

Each entry in passes has the served model, route, request id, token usage and cost. Use the request_id with feedback.

If one pass fails (for example model_unavailable) or returns output that is not valid JSON, the run continues without it. The pass gets an error object and the run's status is completed_with_errors.

These errors stop the whole run and are returned as the HTTP error: insufficient_credits, spend_cap_reached, tier_required, model_not_allowed, model_not_found, content_blocked, key_suspended.

Cost

A run makes 1 intake pass, one pass per reviewer, and 1 aggregation pass (skipped when there are no findings). Each pass is billed like a chat completion on the model that served it; there is no workflow fee. The material is sent in full to the intake, every reviewer and the aggregator, so input tokens grow with the number of reviewers. The default aggregation model, auto-best, uses the large tier, which is usually the most expensive pass. Set aggregate.model to auto to lower cost.

Example with a custom config

from utmost import Utmost

client = Utmost()

run = client.workflows.security_review(
    template="dependency-audit",
    config="""
reviewers:
  - name: known-vulns
    model: auto-fast
    persona: Find dependency versions with publicly known vulnerabilities. Name the fixed version.
min_confidence: 0.4
""",
    input={"content": open("package-lock.json").read(), "filename": "package-lock.json"},
)
for f in run["findings"]:
    print(f["severity"], f["title"], f["location"])
print("cost:", run["cost"]["total_usd"])