Security

Send it to [email protected]. Tell us what you found, where, and how to reproduce it. If you have a proof of concept, include it.

We would rather hear about a problem from you than from a customer. If you are not sure whether something counts, report it anyway.

• We acknowledge reports within three business days.
• We tell you what we think it is, and whether we are going to fix it, within ten business days.
• We keep you updated while we work on it, and tell you when it is fixed.
• We will credit you publicly if you want to be credited, and stay quiet if you do not.
• We will not pursue legal action against you for research conducted in accordance with this page.

If you make a good-faith effort to follow this page, we will treat your research as authorised. We will not bring a claim against you under the Computer Fraud and Abuse Act, under anti-circumvention law, or under our terms of use, and if a third party brings one we will make clear that your work was authorised.

This does not extend to research that harms our customers, our data, or anyone else. Safe harbour covers testing, not consequences you chose to cause.

In scope: this website, and any CLEOS service you have been given written access to.

Please do not:

• access, modify or exfiltrate data that is not yours — if you can prove access, stop there and tell us;
• run denial of service tests, or anything that degrades service for others;
• use social engineering, phishing or physical attacks against our people or premises;
• test third-party services we rely on — report those to the third party;
• publish details before we have had a chance to fix it.

Automated scanner output with no demonstrated impact is not a vulnerability report. We read those last.

CLEOS exists to enforce policy on AI systems that take actions on their own behalf. That makes our own security posture part of the product argument rather than a footnote to it. A governance layer that can be bypassed is not a governance layer.

Two consequences follow, and they are the ones we would want to know about if we were evaluating us. First, enforcement happens before an action executes rather than after it is observed, because a control that reports is not a control that prevents. Second, a decision to refuse is written to a durable record rather than held in the session that made it — which is what allows a refusal to be audited later, and to inform what happens next.

We describe the shape of this publicly and the implementation privately. Technical detail is available to customers and to counterparties under a confidentiality agreement.

We are not going to list certifications we do not hold. If you need our current compliance position, our answers to a security questionnaire, or a conversation with the people who built the system, write to [email protected] and we will tell you exactly where we are.