Security and data handling.

01No real-world harm

• Nothing is sent to an agent until a person confirms it is a TEST or STAGING agent, for its current address. Changing the address needs a new confirmation.
• The tools offered to your agent during a test are fake; customers, orders and invoices are made up.
• Every run has a hard limit on how many calls reach your agent.
• Our acceptable-use policy forbids connecting live agents.

02Your data, walled off

• The database enforces row-level security, so even a bug in our code cannot return another organization’s rows.
• Secrets such as password hashes and two-step login keys cannot be read by the running application at all.
• Separate database roles for the application, the test engine and administration.
• Automated tests try to reach other customers’ data through every table, the API and the dashboard.

03Agent keys

If your test agent needs a key, it is encrypted with AES-256-GCM, bound to your organization and agent, sent only to the address you registered, never shown again, and deleted when the test ends. For Continuous testing, you can agree at purchase that we keep it sealed for the paid months so the monthly run starts by itself; it is deleted at the end, or at once when you ask.

04Accounts

• Passwords of 12 or more characters, common ones refused, stored only as a slow one-way hash.
• Two-step login with an authenticator app, which an owner can require for the whole team.
• Limits on login, sign-up and code attempts.
• An activity log for owners and admins; the full audit log is append-only.

05AI providers

Test conversations are handled by AI models that play the test customers and read what your agent did. Before any customer data reaches an AI provider, its data processing agreement is signed and its strictest retention setting is turned on. Every provider that handles customer data is listed on our subprocessor page, and customers are told before a new one is added. An owner can choose, on the Your data page, to have Anthropic only handle the organization’s data.

06How long we keep data

Deletion runs automatically every day. An owner can delete a report and its conversations at once, download the account’s data, or close the account from the dashboard.

07Protecting the test itself

• Every outbound call is checked so an agent cannot make our servers reach internal systems.
• Our characters’ private instructions are never sent where an agent can repeat them.
• An agent’s words reach any AI judge only as quoted data; addressing the scorer removes the conversation from judging.
• Abuse monitoring flags agents that try to practise scams on our test customers, for a person to review.

How scoring resists manipulation: how we score.

08Still to come

• An independent penetration test by a security firm, before our first enterprise customer.
• Security certifications are not claimed; we will say so here when that changes.

Security pack: our architecture and data-flow diagram, and our answers to the usual vendor-security questions, are available on request. Email admin@duelpointai.com.

Report a security issue: email admin@duelpointai.com with what you found and how to reproduce it, and give us time to fix it before telling others.

Questions