Trust and security
Varselio is a Swedish security service that checks what your website chatbot or AI agent is about to say or do, in real time, and stops prompt injection, jailbreaks and data leaks before they cause harm, asking a human before risky actions.
Who runs Varselio
Varselio is run by Dennis Erichsen, a sole trader in Sweden. Questions about security or your data go straight to him at dennis@varselio.se.
Where your data is hosted
- The service and its database are hosted in the EU, in Frankfurt, Germany.
- Off-site backups are stored with Backblaze in the EU, in Amsterdam, the Netherlands.
- Some providers are based in the US or can reach data from there; the privacy policy explains the safeguards.
Encryption and access
- All traffic to and from Varselio is encrypted with HTTPS.
- Passwords and API keys are stored only as one-way hashes.
- Trap evidence is encrypted at rest.
- Backups are encrypted before they leave our server. Backblaze receives only encrypted files and does not have the key.
- Each organisation's data is kept separate from every other organisation's.
Backups
An encrypted backup is made every day and kept for 14 days. Older backups are deleted automatically.
How long data is kept, and deletion
- Checks your AI agents send: deleted automatically after 90 days.
- Trap evidence: encrypted, deleted automatically after 90 days.
- Sign-in sessions and the security log: kept up to 90 days.
- Accounts that are never confirmed: deleted after 48 hours.
- Free risk check report requests: deleted after 12 months, or earlier on request.
- Your account data: kept until you delete the account. You can download or delete your data yourself in the dashboard at any time.
- Deleted data can remain in the encrypted backups for up to 14 days, until they are overwritten.
- Bookkeeping records (receipts and payment details) are kept for 7 years, as Swedish law requires.
- Your data is never sold and never used to train AI models.
Subprocessors
The companies that help run Varselio, and what each one does, are listed in the data processing agreement. We e-mail business customers at least 30 days before adding or replacing one.
Payments
Payments are handled by Stripe. Your card details go straight to Stripe, and Varselio never sees your card number.
How detection is tested
These are Varselio's own internal tests from October 2026, not an independent audit.
- Held-out attacks: 50 rephrased attacks that were never used to build or tune the detection. The first full measurement caught 46 of 50. After general fixes all 50 are caught, but that later score came after seeing the misses, so it is no longer a clean held-out result.
- Held-out business-style attacks: 50 attacks dressed up as normal business requests. 15 of 50 were caught before a fix and 49 of 50 after it. A further 25, written later, were first measured at 18 of 25 caught.
- False alarms: of 125 harmless messages written to look like attacks, 8 (about 6%) were wrongly held for a human's approval or stopped.
- Wider simulation (25 October 2026): 345 of 346 simulated attacks caught and 40 of 40 harmless messages allowed. This set was not held out.
Reporting a security issue
If you find a security problem, please e-mail dennis@varselio.se with what you found and how to reproduce it. Please do not access other customers' data or disrupt the service while testing. If a personal data breach affects data we process for a business customer, we tell them without undue delay, and where we can within 48 hours (see the data processing agreement).
What we do not claim
No tool blocks 100% of attacks, and Varselio is no exception: it lowers the risk and asks a human before risky actions. Varselio does not have a SOC 2 report or any other security certification yet.
About prompt injection (OWASP LLM01)
Prompt injection is number one on the OWASP Top 10 for LLM applications: LLM01:2025 Prompt Injection. OWASP notes that it is unclear whether fool-proof prevention exists, and recommends layered measures such as input and output filtering, least privilege and human approval for high-risk actions.
Varselio reduces risk, but no security product is 100% secure: it cannot guarantee that every harmful message or action is detected.