AI companies should not be a black box. We promise to publish regular transparency reports that make it possible for anyone — a user, a researcher, a journalist — to check our work. This is the first one.
It covers the period from December 21, 2025 (when Invariant launched) through January 13, 2026. Future reports will run on a regular cadence.
Account actions
In this period we received approximately 460 abuse reports across Pluto and the early Airo previews. Of those, 312 led to an automated action (warning, rate-limit, or suspension) and 41 led to an account termination after human review.
Of the accounts we actioned automatically, 24 appealed. We reviewed every appeal with a human on the team. We reversed 9 of them, partially restored 3, and upheld the rest. When we got it wrong, we said so to the affected user — not in a templated form letter, in an actual email signed by a person.
Refusals and safety
Pluto refused to answer roughly 0.6% of the messages it received. About a third of those refusals were for hard-no categories (weapons of mass harm, sexual content involving minors, targeted harassment, credential theft).
The other two-thirds were closer to judgment calls. Our internal review found that around 14% of those judgment-call refusals were too cautious — Pluto refused something it should have answered with appropriate caveats. We've adjusted the model's behavior in Pluto V2 to reduce that rate, and we'll keep tracking it.
Government and legal requests
We received 0 government requests for user data in this period. We received 0 court orders. We received 0 national-security letters. If and when we do, we'll say so, with the same level of detail we can legally provide.
Our policy is to push back on requests that are overly broad, unspecific, or appear to be fishing expeditions. We will notify affected users unless legally prohibited from doing so.
Security incidents
We had 0 confirmed security incidents involving user data in this period. We had two minor outages (totaling 47 minutes of degraded service) which are documented on our status page.
Things we got wrong
On December 28, a deployment regression caused Pluto to truncate long replies for about 90 minutes. We rolled back, posted an incident report, and added a regression test.
On January 4, an automated abuse system flagged a small number of legitimate study queries as policy violations. The accounts were not suspended, but the affected users saw a confusing warning. We retrained the classifier, removed the warnings from those accounts, and emailed each affected user.
Why we publish this
Transparency reports are easy to write when you have nothing embarrassing to admit. They're meaningful when you do. We'd rather be honest about the messy parts of the first month than wait for a cleaner narrative.
If you have questions, or you think we missed something that should be in here, write to hello@invariant.ai.