Privacy Policy

This Privacy Policy explains how AdversariaLLM ("AdversariaLLM", "we", "us") collects, uses, discloses, and protects information when you use our website, hosted inference platform, developer API, and related content (the "Services"). AdversariaLLM is a platform for authorized security research — it serves open-weight language models specialized for offensive-security, reverse-engineering, and malware-analysis work. Because of that focus, this policy also describes how we handle uploaded files (including potentially malicious samples) and the geographic and export-compliance screening we apply.

This policy is written to be true to how the Services actually work. Please read it alongside our Terms of Service and Acceptable Use requirements, which govern the authorized-use conditions referenced below.

1. Scope

This policy applies to:

• Visitors to our public content (the Research, Field Notes, Learn, and Resources sections, the model catalog, pricing, and documentation);
• People who join our launch waitlist before the Services are generally available;
• Registered account holders who sign in, chat with models, upload files, and use the API.

It does not describe the internal data practices of third parties whose sites you reach through links we publish, or the practices of the identity, payment, hosting, and model-serving providers we rely on beyond the role each plays for us (see Section 9).

2. Information we collect

2.1 Information you provide directly

Account and identity data. An email address is required to create an account. We store your email address, a display name, and — if you choose to set it — free-text "custom instructions" you want applied by default to new conversations. You may sign in through a passwordless email "magic link", through a password (if you set one), or through a third-party identity provider (Google or GitHub). When you sign in with a provider, we store the provider's stable internal account identifier and a snapshot of the email address the provider asserted at the time you linked it; we do not treat that email as a way to silently merge accounts.

Authentication credentials. If you set a password, we store only a salted, hashed verifier — never the password itself. If you enable multi-factor authentication, we store the authenticator's public key and related non-secret metadata and backup-code hashes — not any biometric or private-key material, which never leaves your device.

Conversation content. When you chat with a model, we store your messages, the model's responses, any per-conversation system prompt you edit, and any conversation titles. This content is retained so you can return to your conversations, branch or continue them, and share them if you choose.

Uploaded files. If you attach a file (an image, a document, or a binary sample such as an executable, archive, or packet capture), we store the file, its size, MIME type, original filename, and a SHA-256 hash. For text-bearing documents we may store extracted text so a model can analyze it. See Section 6 for how we handle files safely.

Sharing. If you create a share link for a conversation, we store a frozen snapshot of the shared content and a share token. You control whether a share exists, whether it includes your custom instructions, and when it is revoked.

Payment information. If you purchase a subscription or prepaid credits, our payment processor collects and processes your payment-instrument details. We do not store full card numbers or equivalent secrets. We store records the processor returns to us — subscription and invoice status, amounts (in USD), a processor-side reference identifier, and the billing country the processor asserts for you. If you pay with cryptocurrency for prepaid credits, we store the on-chain transaction identifier and confirmation status; we do not collect wallet identity beyond what a confirmed payment reveals.

Waitlist. Before the Services open, you may submit an email address to be notified at launch. We store that address (in a normalized form to avoid duplicate notices) and which page captured it. A waitlist entry is not an account and grants no access; it is used only to send a launch notice.

Support and other communications. If you contact us, we retain what you send us.

2.2 Information we collect automatically

Session and device context. When you sign in, we record the session, its creation and expiry time, and — so you can recognize and revoke your own sessions — the IP address, user-agent string, and inferred country associated with that sign-in, plus a "last seen" time. We similarly record a coarse source IP for sign-in-link and OAuth requests, used only for abuse forensics and never as an authorization signal.

Usage and billing metadata. For each generation we record operational metadata: which model and version answered, token counts, computed cost, timing, completion status, and references linking the request to the resulting messages. This metadata does not duplicate your prompt or output content.

API-key usage. Each API key is stored only as a hash plus a short display prefix; we never store the full key value. We record the key's name, scopes, creation time, optional expiry, and when it was last used.

Readership analytics (privacy-preserving). For our published content (Research, Field Notes, Learn, and Resources), we count readership in a way that is designed not to identify readers. We do not store a raw IP address, user-agent, or user identifier in our analytics. Instead we store a keyed one-way digest whose only purpose is to avoid double-counting the same reader within a 24-hour window: a signed-in reader is keyed by their opaque session identifier; an anonymous reader is keyed by a combination of address, user-agent, and the calendar day, which rotates daily and cannot be reversed into a person. Automated crawlers are filtered out before anything is recorded.

Cloudflare Web Analytics. On production pages we load Cloudflare's privacy-first, cookieless analytics beacon to measure aggregate traffic (volume, rough geography, distinct-visitor counts). It sets no cookie and builds no cross-site profile. It is disabled outside production.

2.3 Compliance and geographic data

To meet legal and export-control obligations, we record access decisions at sign-up, sign-in, and certain other checkpoints. These records include an inferred country and subdivision, the signal that produced it (payment-provider assertion, edge-geolocation, or a declared value), the subject (your user identifier once you have an account, otherwise a source IP), the decision reached, and supporting evidence. This evidence never contains prompt or output content. We may also screen against denied-party and sanctions lists as required by law. See Section 5.

3. How we use information

We use information to:

• Provide the Services — authenticate you, run your conversations, serve model responses, store and retrieve your chats and files, and enable sharing you initiate;
• Meter and bill usage — reserve and settle credits, produce usage records, process subscriptions and prepaid-credit purchases, and maintain accurate financial records;
• Keep the platform secure — verify email reachability, detect and prevent abuse and fraud, safely handle uploaded samples, and let you manage your own sessions and keys;
• Comply with law — perform export-control and sanctions screening, honor geographic restrictions, and respond to lawful requests;
• Measure and improve — understand aggregate readership and traffic (using the privacy-preserving methods in Section 2.2) and improve content and features;
• Communicate with you — send transactional email (sign-in links, verification, receipts, security and account notices) and, if you asked to be notified, a launch announcement.

We do not sell your personal information, and we do not use it for third-party advertising or cross-site behavioral profiling.

4. Model training and your content

AdversariaLLM serves third-party, open-weight models. We do not use your prompts, conversations, uploaded files, or outputs to train or fine-tune models. Your content is processed to produce the responses you request and to operate the Services described above.

Separately, for the content we publish, we emit a machine-readable Content-Signal of search=yes, ai-input=yes, ai-train=no — our stated position that third parties may index our published articles but should not use them as AI training data. That signal governs our own published content, not your account data.

5. Geographic and export-control screening

Because AdversariaLLM distributes security-research model capabilities, access may be restricted or denied based on jurisdiction and applicable export-control and sanctions law. To do this we infer your location from the strongest available signal (a payment-provider assertion where present, otherwise edge geolocation or a value you declare) and evaluate it against versioned rules. Some models carry export classifications that further restrict where they may be served.

Every denial is recorded with the rule version that decided it, so a decision can be reviewed and appealed. If you believe you were incorrectly blocked (for example, while traveling or behind a VPN), you may request review; a staff override, when granted, is time-limited and audited.

6. How we handle uploaded files, including malware samples

Uploaded files are the highest-risk surface on the platform, and we handle them accordingly:

• A file is quarantined before its contents are trusted. We verify its type against its contents and run an antivirus scan; only files that pass are made downloadable.
• We never execute, run, unpack, or decompress your uploaded samples on our servers as part of ordinary analysis. Binary samples are stored and served only as attachments and are analyzed statically — reading facts out of the bytes at rest (format headers, section entropy, strings, indicators of compromise, capture summaries).
• Any dynamic ("detonation") analysis is off by default, gated, and requires your explicit per-sample consent, which we record along with the version of the consent language you accepted.
• Files and their analysis reports are scoped to you (and your workspace) and are not accessible to other users.

You are responsible for having the rights and authorization to upload and analyze any sample you submit. Do not upload personal data of others that you are not authorized to process.

7. How we share information

We share information only as follows:

• Service providers / subprocessors. With the vendors that host and operate the Services on our behalf, each limited to its function (see Section 9).
• At your direction. When you create a share link or otherwise choose to make content available.
• Payment processing. With our payment processor(s) to complete transactions you initiate.
• Legal and safety. When required by law, legal process, or a lawful government request, or where we believe disclosure is necessary to protect the rights, safety, or property of AdversariaLLM, our users, or the public — including to enforce our terms and prevent abuse.
• Business transfers. In connection with a merger, acquisition, financing, or sale of assets, subject to this policy.

We do not sell personal information.

8. Cookies and similar technologies

We use a small number of strictly necessary cookies:

• A session cookie that keeps you signed in. It is HttpOnly (not readable by page scripts), SameSite=Lax, and marked Secure in production.
• A short-lived sign-in flow cookie used to protect third-party sign-in against cross-site request forgery.

We do not use advertising or cross-site tracking cookies. Our Cloudflare Web Analytics beacon is cookieless. Because we rely only on strictly necessary cookies and privacy-preserving, cookieless measurement, we do not present a tracking-consent banner.

9. Service providers (subprocessors)

We rely on the following categories of providers, each processing only what its function requires:

• Application hosting — runs the website, API, and database.
• Content delivery, edge security, and analytics — Cloudflare, for CDN, edge geolocation used in screening, and cookieless traffic analytics.
• Model serving / GPU compute — the infrastructure that runs the open-weight models; prompts and outputs are processed here to generate responses and are not used to train models.
• Payment processing — our card processor for subscriptions, and a cryptocurrency processor for prepaid-credit purchases where offered.
• Email delivery — the provider that sends transactional email (sign-in links, verification, receipts, notices).
• Identity providers — Google and GitHub, when you choose to sign in through them.

A current list of subprocessors is available on request.

10. Data retention and deletion

We keep information for as long as your account is active and as needed to provide the Services, then as required for the purposes below.

Account closure. You can close your account from your settings. When you do, we anonymize and revoke rather than hard-delete: your email address is replaced with a non-identifying tombstone, your display name and custom instructions are cleared, your credentials are dropped, and all sessions are revoked.

Records we retain after closure. For legal, accounting, audit, and anti-fraud reasons, certain records are append-only and are not deleted when you close your account — in particular financial records (credit ledger, usage and generation metadata, invoices, payment events) and compliance/audit records (access decisions, security events). These are keyed to your internal account identifier after your identifying details have been scrubbed.

We retain each category of information only for as long as needed for the purposes described above and as required by law, after which it is deleted or anonymized.

11. Your rights and choices

Depending on where you live, you may have rights to access, correct, delete, port, or restrict the processing of your personal information, and to object to certain processing. You can already:

• View and revoke your active sessions;
• Create, scope, and revoke API keys;
• Edit or clear your display name and custom instructions;
• Delete individual conversations;
• Close your account (which anonymizes and revokes it, subject to Section 10);
• Unsubscribe from the launch waitlist and from non-transactional email.

To exercise rights not yet self-service, contact us at privacy@adversariallm.ai. We will verify your request against your account before acting on it. We will not discriminate against you for exercising your rights.

12. Security

We protect information with measures including: hashing of passwords, API keys, and sign-in tokens (never stored in the clear); HttpOnly, SameSite=Lax session cookies; single-use, time-boxed, keyed sign-in links; CSRF-protected third-party sign-in; multi-factor authentication support; quarantine-and-scan handling of uploads; and a rule that prompt and output content is never written to operational logs. No system is perfectly secure, and we cannot guarantee absolute security.

13. Children

The Services are not directed to children and are intended for adults engaged in authorized security research. You must meet the minimum age stated in our Terms of Service to use the Services, and we do not knowingly collect personal information from anyone below that age.

14. International users and data transfers

We operate the Services and rely on providers that may process information in countries other than the one you live in. Where required, transfers are made under an appropriate safeguard. Access from some jurisdictions is restricted for the export-control and sanctions reasons described in Section 5.

15. Changes to this policy

We may update this policy from time to time. When we do, we will revise the "Last updated" date the app stamps at the top and, for material changes, provide additional notice as required.

16. Contact

Questions about this policy or our privacy practices can be sent to privacy@adversariallm.ai. If you are in a region that requires it, we will designate a representative and/or data protection officer as needed.