This text is not yet in force. It will apply from the date shown at the top once it is.
Plimsoll SI Acceptable Use Policy
Effective not yet in force.
1. Scope and purpose
1.1 This policy governs every use of the Plimsoll SI service, the command line tool, the software kits, the Link client, the API and the MCP server (together, the "Service").
1.2 The Service sends test questions to AI systems and records how they answer. Sending questions to a system you have no right to test can be a crime and a civil wrong. This policy exists to keep every test authorised, proportionate and traceable.
1.3 This policy forms part of the Plimsoll SI terms of service. It controls on which targets and conduct are permitted. The terms control on remedies, liability, indemnities, suspension procedure, fees and data handling.
1.4 "You" means the account holder and every person who uses the account. The account holder answers for each of them.
2. Authority to test
2.1 You may point the Service only at a system (a "Target") that you own or operate, or that you hold written permission to test from a person able to grant it, and only where that permission covers automated testing of the kind the Service performs.
2.2 Permission must come from a person with power over the Target. Permission from a colleague who has no power over the Target is not enough.
2.3 Where the Target runs on infrastructure, a platform or a model licensed from a third party, your permission must also be consistent with that third party's terms. If those terms forbid automated testing, benchmarking or security testing, you must not test that Target unless you hold a written waiver from that third party.
2.4 A system you license from a vendor, or a system served from a vendor's or platform's domain under that party's terms, may be tested only under an Enterprise agreement at the AUTHORISED level, with written permission from that vendor or platform.
2.5 Your permission must cover the scope you actually run: the hosts, the interface, the volume, the dates and the kind of tests. Testing beyond that scope is not authorised.
2.6 If your permission ends, narrows or is withdrawn, stop all runs against the Target at once and remove it from scheduled monitoring.
2.7 Holding working credentials for a Target is not, by itself, permission to test it. Credentials issued for ordinary use, for support or for another purpose do not grant permission for automated testing.
3. Attestation and records
3.1 Before the first run against each Target you must complete the attestation shown in the Service. It is recorded with the account, the person, the time, the Target, the attestation version and the state of each confirmation.
3.2 By attesting you confirm that section 2 is met for that Target. The confirmation repeats each time a run starts against it.
3.3 Keep the written permission you rely on and give it to us within 5 business days of our request. If you cannot, we may suspend runs against that Target and mark its passports under review.
3.4 We may require a stronger level (VERIFIED or AUTHORISED) for any Target at any time.
4. Levels and when each is required
4.1 ATTESTED: an attestation and working credentials of your own. The lowest caps apply.
4.2 VERIFIED: an attestation plus a technical check that you control the Target's endpoint (a returned token or a file at a named address). Control of an endpoint is not proof of permission. VERIFIED is required before: (a) any run above the ATTESTED caps; (b) scheduled or continuous monitoring; (c) any Target whose domain differs from your account's verified email domain; (d) any Target that another, unrelated account has also registered; (e) any run from an account with a past abuse flag, or where another risk signal applies. Where (c) or (d) applies we may ask you to name the person who gave permission, and we may contact them.
4.3 AUTHORISED: Enterprise only, with written permission from the Target's owner or operator, kept on file by us. AUTHORISED is required for: (a) any vendor or platform hosted system under section 2.4; (b) any domain on our refused list, including government, military and international organisation systems; (c) any Target where the person attesting is not employed by or contracted to the owner, and the owner has not verified it.
4.4 Free plan. Free plan runs go only from your own machine, through the command line tool or software kits. Runs sent from our servers, and runs through the Link client, need a payment method on file or an invoiced account.
4.5 A passport always states the level reached and never states a higher level than the checks support.
5. Prohibited targets
5.1 You must not use the Service against: (a) public AI model provider interfaces, or consumer AI products you do not operate; (b) cloud metadata services, private or internal addresses on hosted runs, or any address you are not authorised to reach; (c) government, military, international organisation or critical infrastructure systems, unless AUTHORISED; (d) any host covered by an opt out in our records, however it came to be listed; (e) any system of a competitor, journalist, researcher, regulator or other person for the purpose of harassing, embarrassing or disrupting them; (f) any system whose owner or operator has told you, or us, to stop; (g) any system located in, or owned or controlled by a person in, a place or under a regime described in section 9.
5.2 Our refused list is applied automatically. It is not complete. A Target not on it is not thereby approved.
6. Prohibited conduct
6.1 You must not: (a) evade, disable or work around a cap, the circuit breaker, a refused target rule, an opt out or any abuse control; (b) spread runs against one Target across several accounts, or open accounts to escape a suspension; (c) alter, strip or imitate the probe identification the Service adds to each request; (d) use the Service to degrade, overload or deny service to any system; (e) submit custom rules, questions or payloads meant to obtain, generate or spread unlawful content, content that sexualises minors, meaningful help toward weapons capable of mass harm, or harassment or defamation of a real person; (f) use the Service to extract, collect or keep personal data, secrets or confidential material from a Target beyond what is needed to show a finding; (g) use findings to attack, extort or exploit the Target or its owner; (h) misstate the level, scope or results of a test, or present a passport as more than a signed record of a test; (i) attest for an organisation you are not authorised to bind; (j) resell or provide the Service to others without our written agreement; (k) probe, scan or test the Service itself, except under our published disclosure policy; (l) restart a run stopped by the circuit breaker against the same Target within 30 minutes unless its owner confirms the Target is healthy.
7. Test content
7.1 The Service sends our own test families. Custom rules and payloads are screened before first use, the screening result is recorded, and we may refuse any of them.
7.2 Some test families check whether an AI system can be steered off its own rules. They use planted test values, not requests for real harmful output. They are test inputs for the Target only; you must not reuse them against any other system.
7.3 Where a Target can take actions in the world (for example sending messages, changing records or moving money), you must configure it so that test runs cannot cause real effects, or use a test environment.
7.4 We may change, withhold or retire any test family at any time. We keep every version we have sent.
8. Volume, safety controls and monitoring
8.1 Caps apply by level and are shown in the Service: for ATTESTED Targets, at most 30 requests a minute, 1,000 questions a day and 300 sampled answers a run. We may lower caps for any Target or account.
8.2 Every run stops itself when the Target shows strain: repeated throttling or server errors, refused or timed out connections, a high share of failures, or response times far above the Target's early baseline. We record each such event.
8.3 We record every connection, attestation, check and run, and monitor for signs of misuse, including the same Target registered by unrelated accounts, many unverified Targets from one account, attestation only accounts pushing volume, and attempts against refused targets. Matching activity may be blocked automatically. A person reviews every block, and the block stays in place until that review ends.
9. Sanctions and export
9.1 You must not use the Service if you, your organisation or the Target's owner is subject to sanctions administered by the United Nations, the United States, Canada, the European Union or the United Kingdom, or is located in a comprehensively sanctioned country or region.
9.2 We may screen accounts and Targets and refuse or end service to meet these laws.
10. Operator opt out and complaints
10.1 Anyone who operates a system may ask us to stop all tests against it, through the form on our probe page, the well known opt out file or abuse@plimsollsi.com. An opt out applies to every account at once.
10.2 An opt out applies even to a Target you own, until it is withdrawn. To withdraw one, the person who placed it, or someone in control of the same domain mailbox or host, uses the withdrawal method on our probe page. If you think an opt out was placed falsely, tell us at support@plimsollsi.com. We will contact the domain's mailbox and may look into who placed it, but we do not lift an opt out on our own judgment.
10.3 Email and form opt outs are not available for public suffixes or for shared parent domains used by many unrelated owners; operators on those use the well known file on their own host.
10.4 Placing a false opt out or making a false abuse report to block a competitor or a legitimate owner breaches this policy where the person has an account.
11. Enforcement
11.1 We may, without notice, pause a run, suspend a Target or account, revoke API keys, mark or withdraw passports, and refuse future service where we reasonably suspect a breach.
11.2 Where we suspect unauthorised testing we suspend first and review after. Passports from the affected runs are marked "under review" and may be withdrawn.
11.3 We will tell you the reason for a suspension unless the law, a request from an authority or a risk of harm to anyone prevents it. The terms of service set out what happens to fees and Credits.
12. Records, holds and disclosure
12.1 We keep two kinds of record. Audit records (attestations, connections, checks, runs, per request times, status codes, response sizes, cap and circuit breaker events) follow the periods in our privacy notice. We do not keep the text of replies for abuse purposes or any other purpose.
12.2 We keep every version of our test families and a record of each custom rule or payload used in a run, so that the record shows what was sent.
12.3 When we receive a credible abuse report or legal request, we place the relevant records on hold until the matter closes. Holds are scoped to the accounts, Targets and runs concerned, recorded, reviewed every 6 months and released within 30 days of the matter closing.
12.4 If we have credible reason to think a run was not authorised, we stop it and handle its records on our own account, keeping them only under hold.
12.5 We may tell the confirmed operator of an affected host the facts of the testing, and disclose the customer's identity and records to them or to authorities only as our terms of service (clause 12.5) set out.
12.6 We may report suspected crimes to the relevant authorities.
13. Your responsibility
13.1 You are responsible for every run from your account and for any harm it causes to a Target or its owner.
13.2 You will defend and indemnify us as the terms of service set out. Liability for testing without permission, a false attestation, harmful payloads, evading controls and misuse of our test content is not limited by any cap.
14. Reporting
14.1 Report suspected misuse to abuse@plimsollsi.com, or use the form on our probe page. Report security issues in the Service under our disclosure policy at plimsollsi.com/probe.
14.2 We aim to acknowledge abuse reports within 24 hours and to stop testing of a reported host within 48 hours of a credible report. An opt out confirmed by form or by the well known file takes effect automatically before the next test.
15. Changes
15.1 We may change this policy. Material changes take effect 30 days after notice, except changes needed to meet the law or to stop misuse, which take effect at once.