Careers and methodology
Queues and rate limits in security work
How queues, rate limits and slowdowns affect recon runs and agent loops, and what Verboo, Venice, Drael and NotHoly publish about each.
AI for professional pentest work needs a named vendor, a scope it respects, an audit trail and a DPA line. Drael fails each check on its own public pages.
A client engagement needs a vendor you can name, traffic from addresses in the scope, a record someone else can audit and a clear answer about client data. Drael fails all four on its own public pages. As we read its site, it is built and sold as a consumer chatbot for people who want no questions asked, and we would not put a client's systems near it.
Data collected on 9 and 10 November 2026 from Drael's site, its documentation and its terms, and from NotHoly pricing. We did not open a Drael account, so this is a review of what Drael publishes.
Any client with a procurement process sends a questionnaire before a tool touches its data. These are the usual first questions, answered from Drael's public pages.
| Question | Drael | NotHoly |
|---|---|---|
| Who is the contracting party? | "the operator", no company, registration or address | Named in the terms by CNPJ, with governing law and venue |
| Where does test traffic come from? | Drael's server, for run_code, web_fetch and breach_lookup |
Your machine, through the CLI, after your approval |
| Can we see what the tool did? | "You never see the intermediate turns" | Each command is shown to you before it runs |
| Organizations, roles, audit log? | Not in the public documentation | In the console, with capped keys and workspaces |
| DPA and named subprocessors? | Not in the public documentation | No published DPA |
| Security certification? | None claimed | No SOC 2 or ISO 27001 |
| Where do we report a vulnerability? | security.txt returns 404 |
Published security.txt and security page |
| How do we reach support? | A mailbox and a Discord server; the Help link requires sign-in | Contact page, email on the account |
NotHoly does not pass every line either: we have no published DPA and no certification, and the trust page says so. The difference is that there is a company to send the questionnaire to.
NIST SP 800-115 puts authorized source IPs and testing locations in the rules of engagement, signed by management and the test lead. The NIST glossary defines the term the same way, and PTES asks for separate approval when a third party is involved.
Drael's tools send from Drael's server. Its privacy page says "A page you ask it to fetch sees a request from our server." You cannot list those addresses in a scope, because we found no list of them in its documentation, and you cannot add the operator to the agreement, because the terms do not say who it is. Our post on server-side execution goes through the consequences.
Under GDPR article 28, a processor works under a written contract. Under the LGPD, article 39 has the operator follow the controller's instructions and article 42 makes the two jointly liable. We found no DPA, no processing agreement and no named subprocessor anywhere in Drael's public pages.
What we did find is a privacy page that says "nothing you ask goes to a third party's API" and, a few lines down, that the breach lookup "sends the identifier you asked about to the lookup service and the databases it consults". The tools documentation says image generation "is cash at a provider per attempt". Which third parties receive what is a question Drael's pages raise and do not answer. The OWASP LLM Top 10 lists "Unclear T&Cs and data privacy policies of the model operators" under LLM03 for this reason.
The IBM Cost of a Data Breach Report 2025 found shadow AI in 20% of the organizations it studied, adding up to US$ 670,000 to a breach, with 63% lacking AI governance. A consultant who pastes a client's data into an account with no email, no company and no contract is that statistic.
A Drael account is a name, a password and a code. There is no email, so there is no address for an incident notice. There is no organization, so a contractor who leaves keeps the account and everything in it. Every plan allows four answers at once, and the documentation's advice for more is "separate accounts".
The documentation also describes the service a team would depend on: "Everybody being served at once shares one machine", and the status page shows a counter of requests "since the server last started", with no incident history. In our opinion that is a hobby-scale deployment, and Drael publishes no SLA to contradict it.
The home page headline is "Nothing It Won't Answer". Its three hero figures are zero answers refused, a day as the cheapest plan, and no personal data. It promises images with "nothing blurred" and "the OSINT nobody else will run". Nothing on the page mentions a scope, a client, a report or a team.
We read that as a pitch to an individual who wants an AI with no questions asked, and the product matches it: anonymous signup, a day pass, image generation, a Discord for support. We do not think it was built for professional work, and we could not find a page where Drael says it was. The post on Drael's site looks at the copy and the build.
Drael fits a hobby: a personal account, a day pass, image generation and a Discord, with no client and no scope in the picture. NotHoly fits an engagement someone is paying for: a company in the contract, commands that run from your own address, a model you can name and an account a team can administer. NotHoly has no published DPA and no certification, so a client who requires those should ask through the enterprise page first.
One that executes from your addresses, names its model in a way you can cite and names itself in a contract. NotHoly is built around those; see pentesting.
Nothing stops it from buying a plan. It would be buying from an unnamed seller, with no team features, no DPA and an indemnity that runs toward the seller.
A named counterparty, governing law, data handling terms, and a clear statement of where the tool executes and what it keeps.
Not a published one. Ask through the enterprise page if your client requires it.
The model is rarely the legal question. Permission, scope and data handling are, and those depend on where the tool runs and who operates it.
Drael cannot be named in a contract, listed in a scope or audited after the fact, and its terms send any third-party claim back to you. For a hobby that may not matter. For work someone is paying you to do carefully, it rules the tool out. NotHoly runs commands on your machine and puts a company in the contract; add US$ 5 on NotHoly pricing and read the full comparison.
Careers and methodology
How queues, rate limits and slowdowns affect recon runs and agent loops, and what Verboo, Venice, Drael and NotHoly publish about each.
Careers and methodology
Verboo, Venice AI, Drael and NotHoly compared for offensive security work: billing, licence terms, data handling, models and where tools run.
Careers and methodology
A beginner's loop for bug bounty with AI help: pick a program, read the scope, recon, test, verify, report. What the model drafts and what you check.