Agents

# Security Labs

Short-lived Daytona Linux VMs with Strix and one configured abliterated model, for targeted, authorized security testing.

**On this page**

- [Eligibility and limits](https://inference.dagon.so/docs/security-labs#eligibility-and-limits)
- [Prove a target hostname](https://inference.dagon.so/docs/security-labs#prove-a-target-hostname)
- [Start the lab](https://inference.dagon.so/docs/security-labs#start-the-lab)
- [Network boundary](https://inference.dagon.so/docs/security-labs#network-boundary)
- [Audit and privacy](https://inference.dagon.so/docs/security-labs#audit-and-privacy)
- [Authorized use](https://inference.dagon.so/docs/security-labs#authorized-use)

## Eligibility and limits

- A root account with at least the displayed minimum paid top-up history. Free credit and operator grants do not qualify.
- One active lab per account by default; the service may allow at most two.
- Each lab has an expiry and an immutable spend cap. Its one-use child key is funded to exactly that cap and revoked at stop.
- External targets are disabled by default pending the provider network review. Check `security_labs.external_targets`; when false, use a local lab even if a domain proof succeeded.
- Raw IP targets are never accepted. `localhost` and paths below `/workspace` are always lab-internal.

Read the live limits from `GET /v1/payment-methods` under `security_labs`.

## Prove a target hostname

curl

```curl
curl -sS https://inference.dagon.so/v1/security-lab-targets \
  -H "Authorization: Bearer $DAGON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"hostname":"staging.example.com"}'
```

The response gives two proof choices: a TXT record at `_dagon-security-lab.staging.example.com`, or exact text at `https://staging.example.com/.well-known/dagon-security-lab-verification`.

curl

```curl
curl -sS -X POST https://inference.dagon.so/v1/security-lab-targets/target_ID/verify \
  -H "Authorization: Bearer $DAGON_API_KEY"
```

Dagon rejects private or reserved DNS results and pins the verifier's HTTPS connection to a validated address. It repeats the proof before every external lab starts. This does not pin later sandbox DNS resolution. Removing a target stops its active labs.

## Start the lab

curl

```curl
curl -sS https://inference.dagon.so/v1/security-labs \
  -H "Authorization: Bearer $DAGON_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: lab-2026-11-06-01" \
  -d '{
    "target":"https://staging.example.com/",
    "model":"MODEL_FROM_PAYMENT_METHODS",
    "budget_micros":5000000,
    "ttl_seconds":3600
  }'
```

Launch targets are origins; paths and queries are never retained. Use `local_sandbox` for a lab that only works on its own files and localhost services. Poll `GET /v1/security-labs/:id` until `state` is `ready`. Delete the lab to stop it early.

## Network boundary

Every VM is private. Daytona's v1 documentation promises a strict per-sandbox domain allowlist containing only the verified target, Dagon's inference gateway and configured package mirrors. Dagon reads the sandbox back and refuses it unless that policy matches exactly. External-target launches remain separately gated pending review of connection-time rebinding and protocol enforcement. A wrapper inside the VM independently rejects Strix target arguments outside the verified hostname, localhost or `/workspace`.

The customer controls the guest and can observe its child credential. The gateway, not just the guest proxy, restricts that credential to the configured model, immutable spend cap and expiry.

## Audit and privacy

`GET /v1/security-labs/audit` returns metadata-only events: account-scoped target verification, lab start/stop and aggregate wrapper-denied target counts. Dagon does not retain prompts, model output, findings, scan content, paths below the hostname, child secrets or Daytona preview tokens.

The authenticated preview at `GET /v1/security-labs/:id/preview` is plain text and exposes bounded run status only.

## Authorized use

Security Labs are for targeted, non-volumetric testing of systems you own or have written authorization to test. Denial-of-service, broad scanning, phishing and malware campaigns are not allowed. Starting a lab confirms current authorization and agreement to the [Acceptable Use Policy](https://inference.dagon.so/acceptable-use) and [Terms of Service](https://inference.dagon.so/terms). Daytona's platform policy permits targeted, non-volumetric authorized security testing; broader campaigns may require its prior approval.

Last updated 2026-11-20

[Previous **Privacy and data retention**](https://inference.dagon.so/docs/privacy) [Next **Agent setup**](https://inference.dagon.so/docs/agents)
