
PEACH
Un cadre d’isolation des locataires
CVE-2026-77332 is an improper output encoding vulnerability in memtomem-stm, a short-term memory proxy gateway with proactive memory surfacing for AI agents. Untrusted long-term and working-memory fields are not properly serialized as inert data before being injected into the surfaced-memory envelope, allowing crafted memory content to introduce Markdown or HTML-compatible structure that can create misleading structural boundaries or visually conceal content from consuming AI agents. The vulnerability affects versions >= 0.1.0 and <= 0.1.33, and was fixed in version 0.1.34, published on July 11, 2026 (GitHub Advisory). It carries a CVSS v4.0 base score of 2.3 (Low), though the GitHub advisory rates its severity as High (Feedly).
The root cause is classified as CWE-116 (Improper Encoding or Escaping of Output): untrusted fields such as memory content, namespace, source file paths, working-memory keys/values, and chunk IDs were rendered directly into the Markdown-compatible <surfaced-memories> envelope without escaping control characters, HTML delimiters, Markdown syntax characters, Unicode bidirectional overrides, or compatibility confusables (GitHub Advisory). An attacker with write access to the memory store (low privilege required) could craft a memory entry containing sequences like \r\n\``html\n&\u202e</surfaced-memories>\x00` to escape the structural boundary of the memory block, inject false structural delimiters, or conceal content from the consuming AI agent (GitHub Commit). The attack requires the attacker to have the ability to write crafted entries into the memory store that will later be surfaced to an AI agent, and the attack complexity is low once that precondition is met. No public proof-of-concept exploit code has been identified.
Successful exploitation can cause a consuming AI agent to misinterpret the memory context by creating false structural boundaries within the surfaced-memory envelope or visually concealing legitimate memory content, potentially enabling prompt injection or logic bypass attacks against the agent (GitHub Advisory). The integrity impact is low — an attacker can manipulate how the agent perceives retrieved data but cannot directly exfiltrate data or cause availability loss. The vulnerability does not address residual model-level risks from ordinary natural-language persuasion, which remains a separate concern outside the scope of this structural serialization flaw (GitHub Commit).
No in-the-wild exploitation has been reported, and no proof-of-concept exploit code is publicly available (Feedly). The EPSS score is 0.0, indicating a very low probability of exploitation in the near term. The vulnerability is not listed in the CISA Known Exploited Vulnerabilities (KEV) catalog. Exploitation requires low privileges (write access to the memory store) and the presence of an AI agent that consumes surfaced memories from the affected gateway.
\r\n\``html\n&\u202e</surfaced-memories>\x00— designed to escape the<surfaced-memories>` XML envelope or inject false structural boundaries.```), HTML tags (e.g., </surfaced-memories>), Unicode bidirectional control characters (e.g., U+202E RIGHT-TO-LEFT OVERRIDE), null bytes (\x00), or CRLF sequences (\r\n) in content, namespace, key, or value fields.Upgrade memtomem-stm to version 0.1.34 or later, which serializes all untrusted fields as inert single-line data by escaping control/bidi characters, HTML delimiters/entities, Markdown delimiters, and Unicode compatibility confusables before injection into the surfaced-memory envelope (GitHub Release). The fix also validates memory IDs against a conservative grammar (^[A-Za-z0-9][A-Za-z0-9._~:/@+%=-]{0,255}$) and pins an explicit untrusted-data preamble to the memory block. As a workaround for those unable to upgrade immediately, disable write-tool surfacing (already the default behavior) and carefully review all retrieved memory context before allowing the agent to take consequential actions (GitHub Advisory).
Source: Ce rapport a été généré à l’aide de l’IA
Évaluation gratuite des vulnérabilités
Évaluez vos pratiques de sécurité cloud dans 9 domaines de sécurité afin d’évaluer votre niveau de risque et d’identifier les lacunes dans vos défenses.
Obtenez une démo personnalisée
"La meilleure expérience utilisateur que j’ai jamais vue, offre une visibilité totale sur les workloads cloud."
"Wiz fournit une interface unique pour voir ce qui se passe dans nos environnements cloud."
"Nous savons que si Wiz identifie quelque chose comme critique, c’est qu’il l’est réellement."