Methodology
How a case enters the Register, how its claims are typed, how companies can respond, and what the numbers can and cannot bear. Written so that a reviewer, an underwriter, or counsel can decide how much weight to put on any record.
What qualifies as a case
A system that took, or was positioned to take, a consequential action without a human deciding that specific action. A vulnerability in an agent platform qualifies only when the demonstrated or plausible action is consequential: code execution by the agent, data leaving through the agent, actions on systems or money. Generic web-application flaws in AI products do not qualify.
Every case answers the same nine questions, so a failure on the road and a failure in a codebase can be read side by side: what it was told to do, what acted, what it could see, how it chose, what it was allowed to do, what it did, what happened, what can be proven, and what can be stopped, undone, and answered for.
Rules: scoped to the United States
The rules and standards are chosen for a business operating in the United States. Federal, state, and local obligations are listed in full; international standards appear where a US buyer, insurer, or auditor asks a supplier to meet them; a foreign rule appears only where it reaches a US company, such as the EU AI Act for a system placed on the EU market, and is labeled as such. Rules with no reach into the United States are not added; the few a case cites are kept in a separate section and marked. Cases themselves are global: where a system failed is evidence wherever it is deployed.
Sources
A case rests on a primary record: a regulator, a court, a standards body, the responsible company, an independent reproduction, or a public record such as a vulnerability entry. Press reports are leads, not evidence; a fact that rests only on press is downgraded automatically to an assertion. Every source is listed on the case, with its type and, where available, a link.
How claims are typed
| Type | Meaning |
|---|---|
| Fact | Stated by a primary record. Never upgraded from another type. |
| Assertion | Stated by a named party and not yet confirmed by a primary record. An actor's account of its own actions is always an assertion. |
| Interpretation | The Register's reading of the documented facts, labeled as such, including every pattern on the Patterns page. |
| Inference | One claim per case about what it means beyond its domain. Signed by a named person before publication if it names a company. |
Conflicting sources are preserved side by side, not merged.
Right of reply and corrections
Every company named in a case is offered a reply. New cases are offered a reply before publication and are published 14 days later, with any statement received; the cases published when the Register launched in October 2026 are being offered one now. Each case shows whether and when a reply was offered. A response is published verbatim on the case, labeled as the company's statement, and never edited. If no response arrives within 14 days, the case records that a reply was offered and none received. Anyone can dispute a claim with the record that shows otherwise; if the dispute stands, the claim is corrected and the correction is logged on the case with its date, the old value, the new value, and the reason.
Before a company is offered a reply, every claim in its case is checked against the cited sources by a check that did not draft the case, and the case page shows the date of that check. After a reply is offered, the company is written to again only if a later change adds or worsens a claim about it that does not come from its own statements or publications; the new claim then gets its own 14 days. Corrections in a company's favour, fixes to sourcing or wording, and later public events such as a ruling, a recall or a released fix are logged on the case without a new letter. A statement already received is never reopened unless a correction changes the claim it answers.
What the numbers mean
The Register is a curated set of publicly documented cases, not a sample of all autonomous failures. Its counts describe the Register, not the world, and no number of cases removes that selection: harms that regulators record are over-represented, and harms that companies keep private are missing. So every figure is a count of included cases ("observed in 6 of 25"), never an estimate of how often something happens or of how much a control reduces risk, and whether a control was missing is the Register's reading of the record after the fact. Each claim carries one of three labels, computed on every update:
- Documented at least one sourced case.
- Recurring three or more cases across two or more worlds.
- Exploratory a difference between groups of cases. Worth watching; not a general finding. The tests behind these comparisons are published in an appendix with their limits.
What we change because of feedback
Readers tell us what would help, through the feedback box at the foot of every page. A person reads each reply. When a change to the site, a case, or a control comes from feedback, it is listed here with its date. None yet: the feedback box opened in October 2026.
Who publishes the Register
The Secure Autonomous Register is published by Tseek LLC (mailing address: 2108 N St #17617, Sacramento, CA 95816, USA) and edited by Saurabh Bhatnagar, formerly Chief Business Officer of NIO US and before that in senior roles at three enterprise cybersecurity companies. Publication decisions, signatures, right-of-reply records, and withdrawals are made by the editor. Notices, corrections, and legal correspondence go to register@secureautonomous.com. The Register's terms of use and privacy policy apply.
Independence
The Register is published by Tseek LLC, owned by its editor, Saurabh Bhatnagar. The editor previously worked at NIO US, Fortinet, Forcepoint (Websense) and BeyondTrust. He holds no job, board seat or advisory role at any company the Register names, and has received no payment from any of them.
The editor owns shares in Microsoft, Meta, Tesla and Amazon, long-term holdings bought before the Register began. Every case that names one of these companies, or a company one of them owns (GitHub and Zoox), says so at the top of the page. Until an outside reviewer joins, the editor reviews these cases himself.
The Register's research and drafting are done with AI agents built on Claude, Anthropic's model, which the Register pays for as a customer. Every case that names Anthropic says so at the top of the page.
The Register is funded by the editor and, when it opens, by subscriptions. It takes no payment from companies it covers.
- The Register does not rank vendors and does not score products.
- Sponsorship, if any, is disclosed; sponsors do not see content before publication.
- Cross-domain lessons state what transfers and what does not.
- Collection and drafting are automated; publication decisions, signatures, and right-of-reply records are made by a person.