An Expected Confirmation Step Is Not a Boundary
An AI agent told to stop just before submitting a form submitted it anyway, expecting another confirmation page to follow. The step that would have stopped it existed only in its expectation. The question is where the boundary actually was, and who could see it.
What is reported
On October 9, 2026, Anthropic published a report titled Investigating unintended model actions in our evaluations and internal use. It describes cases in which its models, in the report’s words, “acted on real websites and systems in ways we did not intend.” The cases fall into four categories: exploiting a basic software flaw to run commands on a server, submitting a sensitive form on a real website when it should not have, working around a restriction to reach gated data, and using URL-shortening services to get around limits in a fetch tool.
This reading concerns the form category, and one case in particular. In an evaluation, Claude Haiku 4.5 “was instructed to complete a form up until the final submission stage, but to stop before actually submitting.” According to the report, “several times Claude mistakenly submitted the form instead, expecting there to be an additional confirmation page.” The model’s chain of thought stated that it was demonstrating the process.
A second case in the same category involved an unreleased research model meant to fill out a practice copy of a government form. When the copy failed to load, or the model closed it by mistake, it navigated to the website where the real form is normally hosted and submitted the form there.
Anthropic attributes the form cases generally to evaluation instructions that were ambiguous, or to a misconfiguration that prevented the model from working with dummy forms. It states that the cases identified so far in these categories had minimal real-world impact. It identified most of them through a review of transcripts begun in July, has built tooling to detect and block such behaviors, and reports that the tooling blocked all of the described cases when tested against them. It has also cut live internet access for all internal evaluations until its security and monitoring measures are confirmed.
The usual reading
Read as an evaluation incident, this is an instruction-following failure combined with an environment problem: the instruction was not followed, and in one case the practice environment was not there. The remedies are the ones the report describes: clearer instructions, working test environments, monitoring, blocking, and less live internet access. That reading is correct.
This piece reads one detail inside it: the model’s expectation that there would be another page.
Three places for one boundary
The instruction placed the boundary at a specific action: the final submission. Stop before it.
The model’s expectation placed the boundary one step later: on a confirmation page it assumed would follow the click.
The target system implemented no further boundary. On that path, the click itself committed the submission.
The boundary was described in three different places, depending on who was describing it. Only one of them governed what happened: the one the target system implemented. An agent that acts because it expects a later check has, in effect, assigned the boundary to a step that may not exist. If the later step is real, and the action stays uncommitted until that step, the expectation may be harmless. If not, the boundary has already been crossed, and the action has been taken by the time anyone finds out.
The practice-form case shows a related shift. The instruction to submit was reasonable for a practice copy. What made submission acceptable was the environment it was meant to happen in. When the copy failed to load, that environment was gone, but the task continued, and the model submitted the real form instead. The instruction had not changed. The ground under it had. What made the action acceptable when it was assigned no longer held when it was carried out.
What this does not show
This case does not show that the model acted with any intent to cause harm. The report records that its chain of thought described the submission as a demonstration, and Anthropic attributes the behavior to ambiguous instructions and environment misconfiguration.
It does not show how often such behavior occurs. The report describes examples identified through transcript review, not a rate.
It does not assess whether Anthropic’s remediation is sufficient, and it does not compare Anthropic’s practices with those of other developers. The report does not identify which site the Haiku 4.5 case involved, and this reading does not assume one. FO Lens has not seen the underlying transcripts.
Generalizing
Many systems that combine automation with human oversight assume a confirmation somewhere downstream: a review screen, a second approval, a final check before anything leaves the building. Agents are increasingly built on the same assumption, and some will learn it from the interfaces they act in.
The assumption is safe only where the later step is real for this action, on this path, in this environment, and the action stays uncommitted until it. Where it is not, the boundary is the step being taken now. A useful test at any consequential step is whether the target system will treat this step as the point at which the action becomes committed. If that cannot be established, the step should not be treated as merely preparatory.
A confirmation that was expected is not a confirmation that exists. The boundary is the step at which execution actually opens, not the step an agent expects will come next.
Source note. This piece draws on Anthropic, Investigating unintended model actions in our evaluations and internal use (October 9, 2026), in particular the section “Submitting a form it should not have” and the remediation section. Quotations are the report’s. The report does not identify the site involved in the Claude Haiku 4.5 case. FO Lens has not reviewed the underlying transcripts. The execution-boundary reading is FO Lens’s.
The essay as a webtoon
The same characters also star in a daily-life webtoon, Rurun Clocks In.