Software agents are repeating the authority mistakes that vehicles, aircraft and plants have already paid for. The Register documents both, so you can see what failed, which control was missing, and what the rules require before you decide what an agent may do without a human approving each step.
Sectors radiate from the center; the further out a case sits, the less a human was deciding. Click a sector or a ring to select it; selections add up. Point at a dot to preview a case, click to open it.
Patterns that recur across vehicles, aircraft, plants, trading systems, and software agents. Each is a count of included cases with its evidence label, recomputed on every update.
The Register cannot tell you this is safe; it can tell you what to verify first. In each of the four direct cases in the brief, at least one of the controls ranked in the brief was missing or failed, and the harm, or in the disclosed cases the demonstrated attack, ran through that gap. The Register has no record of deployments that went well, so it cannot show that these controls are enough, or how much each one reduces risk. If you proceed, treat those controls as conditions to verify in your own system, enforced by the platform rather than by the agent's instructions, and keep a human approval on any action that cannot be undone until you have evidence those conditions hold.
Verify first: least-privilege credential scoped to task, containment and isolation.
Vulnerability entries, recalls, and rules that touch systems acting on their own, tagged with the control each concerns. Posted nightly.
New cases and signals, once a week. A short email when something touches the systems you run. Free; unsubscribe in one click.
Get the weekly emailBuilding agent guardrails? Your AI assistant or agent can query the Register directly, no key needed: https://secureautonomous.com/api/mcp
Cases rest on regulators, courts, the responsible company, or an independent reproduction. Press reports are leads, not evidence.
Each field is labeled fact, assertion, interpretation, or inference, with its source and a confidence. Facts are never upgraded.
Every company named is offered a reply, and each case shows when. For the first cases, published at launch, the offer went out on 14 October 2026, after publication; every case added since waits 14 days for a reply. Responses are published verbatim; corrections are logged with the old and new value.
Incident databases such as the AI Incident Database and the OECD AI Incidents Monitor collect reports. Threat catalogs such as MITRE ATLAS and OWASP's agentic Top 10 list attack techniques. The Register types each claim, maps each case to the control that was missing and the rules that apply, and offers every named company a reply.
Every figure counts included cases and carries its evidence label. The Register describes documented cases, not the world, and says so.
A person reads every answer; changes it leads to are logged in the methodology. To dispute a claim or reply for a named company, use Respond, dispute, or submit.