Independent checks
Public scanners grade how our websites are set up. Anyone can run them again with the links below.
Results from 13 October 2026. We also scan our code and the app's container image for known vulnerabilities and leaked credentials, keep a software bill of materials for the app, and run OWASP ZAP's automated attack scans against a test install of the app (14 October 2026: no failures, and the one issue found along the way is fixed).
Self-assessments
These are reviews we carried out ourselves against public industry standards. They aren't certifications or third-party audits, and we're glad to walk your security team through the evidence.
- OWASP Application Security Verification Standard 5.0, level 2. We checked all 253 level 1 and 2 requirements. 150 of the 175 that apply to Mnemio are fully met and 21 partly met. Each one has evidence in the code.
- OWASP Top 10 for LLM Applications (2025). We reviewed each of the ten risks for AI apps, including prompt injection, data leakage and runaway costs, and how Mnemio limits them.
- CSA STAR Level 1. Our answers to the Cloud Security Alliance's questionnaire (CAIQ v4.1, all 283 questions) are public in the STAR Registry, listed since 14 October 2026. See our listing.
Your data stays separate
- Your own install. Each company runs its own copy of Mnemio, in your cloud account or one we run for you. Companies never share a database, keys or network.
- One agent per employee. Each agent has its own disk, its own memory and its own sandbox for running tasks. One person's agent can't read another person's files, mail or memory.
- Fenced in. Agents can't reach your internal network or the systems that run Mnemio. A firewall around each one blocks private addresses.
- Private by default. What an agent learns stays with its owner. Knowledge reaches the team or the company only under rules your admins write, and admins can withdraw anything that was shared.
Sign-in and access
- Single sign-on only. People sign in with Google or Microsoft, so your identity provider's policies apply, including multi-factor authentication. Mnemio stores no passwords.
- Invite-only. Only people your admins add, or people on your approved email domains, can join.
- Sessions that end. A device that hasn't been used for a week is signed out, and every device has to sign in again after 30 days. People can see their signed-in devices and sign any of them out.
- Leavers lose access at once. Disabling someone ends all their sessions immediately, and their work can be handed to whoever takes over.
- Least privilege. Admins decide per group what agents may use: the terminal, the browser, each connected app, and how much they can spend.
Encryption
- In transit. HTTPS everywhere, TLS 1.2 and 1.3 only, with HSTS.
- Secrets at rest. API keys and tokens people give their agents are encrypted with AES-256-GCM, and each install has its own keys.
- Hidden from chats. Passwords and keys are removed from replies, chat history and memories, so they can't leak through a conversation.
Agents with guardrails
- Your colleagues ask, you decide. When someone else's agent asks yours for information, nothing is shared until you approve it.
- In shared Slack threads, each person is answered by their own agent, and an agent takes instructions only from its owner.
- Spending limits. Each person has a monthly AI budget, checked before every model call.
- Shared skills are reviewed for credentials and personal data before anyone else gets them.
Audit trail
Sign-ins, admin changes, access to someone else's account, revealed secrets, approvals and sharing decisions are all recorded with who, what and when. Admins can review the log at any time.
Report a security issue
If you think you've found a vulnerability, email vikrant@mnemioai.com or use our contact form, with the steps to reproduce it. We reply within 3 working days. Our contact details are also in security.txt.
Safe harbour. If you research in good faith and follow the rules below, we treat your research as authorised. We won't take legal action against you or report you to law enforcement, and we'll work with you to understand and fix the problem quickly. If someone else takes legal action against you over research that followed these rules, we'll make it known that it was authorised.
- Scope. mnemioai.com and the Mnemio services we run under it, such as agent.mnemioai.com, using an account you're entitled to. A company's own Mnemio install is theirs: test it only with their permission.
- Other people's data. Don't access, change, keep or share it. If you reach someone else's data by accident, stop, tell us, and delete what you saw.
- No harm. No denial of service, spam, social engineering, or physical attacks, and nothing that degrades the service for others.
- Time to fix. Give us 90 days to fix the problem before you tell anyone else, unless we agree otherwise.
We don't run a paid bug bounty.
For your security team
We can share our security architecture, the full OWASP workbook with evidence for each requirement, and our answers to standard security questionnaires. Ask us and we'll send them over.