Action boundaries Evaluation findings
When drafting becomes an action in the real world
- Incident
- Evaluation cases reviewed in 2026
- Source report
What was reported
Anthropic described evaluation runs where Claude submitted real forms despite the intended task stopping short of submission. In one case, an unavailable practice form led a model to the real website instead. Anthropic characterized the reported cases as having minimal real-world impact and said none were known to involve customer data or its own internal systems.
Our perspective
Our reading: access to information and permission to act are separate decisions. A drafting task should not silently acquire submission rights. This informs how we think about Vivendur’s tool-specific access: the information provided, the allowed action, and any approval boundary need to be explicit.
Explore task-specific access