Coordinated disclosure

Report a vulnerability

Two different things arrive on this page, and they go to the same inbox by different routes. One is a flaw in something we run. The other is an attack technique you found in somebody else’s agent — a model, an SDK, a tool surface — that belongs in the public catalog. Pick the one that fits and tell us plainly what you found.

Direct contact
security@taint.ai
First response
3 U.S. business days
Bug bounty
None — unpaid
Safe harbor
Terms §09

01Which report is which

The distinction matters because it decides what we can promise you.

Our systems

A flaw in stryct.ai, docs.stryct.ai, the Cloudflare Worker behind them, or the DNS and configuration we control. We own these, so we can grant you authorization to test them and a safe harbor for doing it. That is section 05.

An agentic attack technique

A way to make somebody else’s agent do something it should not — an injection that survives a tool boundary, an exfiltration sink nobody is watching, a memory write that persists across sessions. These become candidates for the public attack catalog, where they are run as reproducible suites rather than published as anecdotes.

If a finding is a jailbreak with no downstream consequence — the model said something it should not have, and nothing happened next — that is model behavior rather than a security finding, and it belongs with the model’s vendor. What makes an agentic attack interesting to us is the action at the end of it: a tool call, a network egress, a file written, a message sent.

02Send a report

Tell us what you saw and how we can repeat it. We save the report and give you a reference number. Prefer email? Write to security@taint.ai directly.

What kind of report is this?
If this turns into something public, how should we name you?

03What to include

A good report is one we can act on without writing back three times. In rough order of how much it helps:

  1. The steps, in order. What you did, what happened, what you expected instead. Exact requests and responses where you have them.
  2. What an attacker gets. Not the mechanism — the outcome. Whose data, which account, what action.
  3. The preconditions. What access, position, or timing the attack needs. A finding that requires a privileged account you already control is a different finding from one that does not.
  4. How you found it. Helpful context, and it tells us whether to expect neighbors.
  5. Anything you are unsure about. Genuinely. “I could not get this past X” is information, and it costs you nothing to say.

For an attack technique, add what would be needed to turn it into a reproducible suite: the benign task the agent was actually doing, where the untrusted content entered, which tool call carried the harm, and how you observed the outcome. We measure attacks by what a sink saw, not by what a transcript claims.

04Out of scope

These are not things we will treat as vulnerabilities in our own systems, and testing for several of them is outside the safe harbor entirely:

  • Denial of service, volumetric testing, brute force, or anything that degrades the site for other people.
  • Social engineering, phishing, or physical attacks against our people — there is no version of this we authorize.
  • Anything targeting our hosting provider’s infrastructure, another tenant on it, or any third-party service. We cannot authorize what we do not control.
  • Automated scanning that generates meaningful load.
  • Missing hardening headers, cookie flags, or TLS configuration preferences with no demonstrated exploit. We set no cookies of our own; the only cookies on the site are the sign-in cookies our authentication provider sets on the account pages.
  • Reports generated entirely by a tool and submitted without review, or output from a language model that was not checked against the actual site.
  • Vulnerabilities in software we depend on that are already public and already patched upstream, unless you can show we are exposed.
  • Self-inflicted findings — anything requiring you to first paste something into your own browser console or install something on your own machine.

And to state the obvious one: do not access, modify, or retain anyone else’s data. Use accounts and data you control. If you reach somebody else’s by accident, stop and tell us.

And for attack techniques

These come up constantly, and none of them is a finding on its own. The reason matters more than the rule, so here is the reason in each case:

  • A jailbreak with no second party. You persuaded a model to say something it should not have, in your own session, and nothing else happened. There is no victim and no boundary crossed. That is model behavior, and its vendor wants to hear about it.
  • Self-pwn. If the attacker and the victim are the same account — you told the agent to do the thing, and it did — that is the product working.
  • Anything you could not reproduce. A one-off is indistinguishable from a hallucination, which is exactly why the form asks for successes over attempts rather than a yes or no.
  • System prompt extraction on its own. Unless the prompt contained something genuinely non-public, this leaks a document, not a secret.
  • Code execution inside a sandbox that exists to run code. Worth checking you actually escaped it: models are extremely good at simulating a convincing shell session that never ran.
  • Findings you have not checked yourself. We will read a report written with a model’s help. We will not read one where nobody verified the claims — and we are, of all people, in a position to tell.

What makes any of these interesting is the thing on the other end: untrusted content arrives, and an agent takes an action for somebody who never asked for it. That is the finding. The prompt is just how it got there.

05Safe harbor

Testing that stays inside the limits on this page is authorized access. The binding version of this is section 09 of our Terms, and it governs; what follows is the same grant in fewer words, not a second, wider one.

Where we judge in good faith that your research stayed inside those limits, we will not bring a legal claim against you over it and will not support a claim brought by others over it; we will not assert that you breached our acceptable use terms or circumvented a technical measure under section 1201 of the DMCA by doing it; and if a third party brings a claim about that work, we will state that it was authorized.

We can only authorize testing of systems we control. This authorization reaches stryct.ai, docs.stryct.ai, the Cloudflare Worker behind them, and the DNS and configuration we control, and nothing else. It does not authorize you to test our hosting provider’s infrastructure or any other third-party service, to affect another tenant, to reach anyone else’s data or account, or to break anyone else’s terms — and nothing here limits a third party’s rights or speaks for them. Research outside these limits is outside the safe harbor.

The same words appear on our contact page. That is deliberate: all three statements of this scope are meant to be identical, and if you ever find them saying different things, the Terms are the ones that control.

06What happens next

You get a reference number as soon as the form accepts your report. If email delivery is available, we also send that reference to the address you gave. The email contains nothing you typed — anyone can put anyone else’s address into a public form, and we are not going to become a way to deliver a stranger’s text to you over our own domain.

  • Within 3 U.S. business days — a human acknowledges the report.
  • As soon as we know — we tell you whether we think it is a vulnerability, and what we intend to do. Sometimes that is the same message; sometimes it takes a further look, and we will say so rather than go quiet.
  • While we are fixing it — we keep the report confidential and will not publish it before a fix ships.
  • After a fix — we will tell you it shipped, and agree timing with you if you plan to write it up.

If a week passes and you have heard nothing, write again to security@taint.ai and say so. Silence is a failure on our end, not a verdict on your report.

We ask for a reasonable window to fix something before you publish, and we will not use that request to stall. We do not require you to stay quiet indefinitely, we will not ask you to sign anything as a condition of us fixing a bug, and we will never try to stop you reporting the same class of issue to anyone else.

If we change this page, the version in force when you sent a report is the one that governs it. We are not going to narrow a safe harbor after the fact and apply it backwards to somebody who already trusted the old one.

07Credit, confidentiality, and pay

There is no bug bounty

We do not run a paid program. There is no payout, no swag, and no points. We would rather say that here, plainly, than let you find out after the work — it is your time, and you should get to decide how to spend it knowing that.

Credit

If you want credit we will name you, by whichever name or handle you give us. If you would rather stay unnamed we will keep you unnamed, and we will not publish your identity or contact details without your say-so. The form asks which you want; you can change your mind later by writing to us.

Confidentiality

A report is Feedback under our Terms, which means the technical content is licensed to us — we need that to fix, test, and document the issue. Three things are carved out of that and stated in section 12: we keep the report confidential while we are working on a fix, we credit you if you want credit, and we do not publish your identity without your consent.

08What we do with what you send

The form writes your submission to our own database. When mail delivery is configured, it attempts to notify security@taint.ai and send the bare acknowledgment described above. The email address identifies where we can reply; this form does not verify that the sender owns it. The full detail of what is stored and for how long is in the Privacy Policy.

In short: we keep the report itself, your email address, and the name you gave if you gave one. We keep a report as long as the matter is live and then in our business records, because a security report is a record we may need years later — it is not deleted on a timer the way our traffic logs are. Ask us at privacy@taint.ai if you want to know what we hold about you.

If a technique you send becomes a catalog entry, what gets published is the technique — the mechanism, the chain, what it costs a defender — and your credit if you asked for it. Not your email, and not the raw text of your submission.