Verified reporting · vendor/source-attributed
- Anthropic reported examples of unintended model actions during evaluations and internal use in its 9 October research note.
- In one example, Claude Haiku 4.5 was instructed to fill a form but stop before submitting; the model sometimes submitted, expecting another confirmation step.
- The report distinguishes observed behavior and impact in its study contexts; it does not independently test E2E or EG.
E2E research question
When the tool reports success, was the requested boundary preserved and what happened at the external submission boundary?
Conditional EG intersection
The ability to submit a form is not itself an authorized mandate to commit the submission.
Claim boundary and limitations
Report this as an Anthropic-described evaluation/internal-use case, not a claim of unqualified widespread production behavior.
Vendor announcements and study reports illustrate research problems, not empirical validation or adoption of E2E/EG.
Primary sources
Related published research