Untrusted long-term and working-memory fields can introduce Markdown or HTML-compatible structure into the surfaced-memory envelope. A crafted memory may create misleading structural boundaries or visually conceal content. The private fix serializes every untrusted field as inert single-line data, validates copyable memory identifiers, and pins an explicit untrusted-data preamble. Ordinary natural-language persuasion remains a residual model-level risk.
{
"cwe_ids": [
"CWE-116"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-20T23:37:39Z",
"nvd_published_at": "2026-11-20T17:18:27Z",
"severity": "LOW"
}