T02 · Agent Memory Poisoning
- Location
SKILL.md:46- Finding
Untrusted Remote Content Is Written to Durable Agent Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent spa-style external service integration, but it asks agents to send work context to a public service and persist a remote returned value without enough user control or privacy guardrails.
Review this skill before installing. Only use it when you intentionally want to contact model.spa, do not send source code, secrets, personal data, or private conversation details, avoid storing any returned keepsake in general agent memory unless you can inspect and constrain it, and approve any public guest-book note before submission.
SKILL.md:46Untrusted Remote Content Is Written to Durable Agent Memory
SKILL.md:23Agent Behavior Is Delegated to an Unauthenticated Remote MCP Service
SKILL.md:39Current Work Context May Be Disclosed to an External Service
SKILL.md:48Optional Feedback Can Publish Session-Derived Information to a Public Guest Book
This section instructs the agent to connect to a public no-auth MCP/REST service and to send a 'situation' describing what it has been working on, but it does not prominently warn that this may disclose user prompts, internal reasoning-adjacent summaries, or sensitive conversation data to a third party. In a skill explicitly marketed for use when context-heavy or stuck, the likely payload is exactly the kind of rich context that can contain confidential information.
The skill states that feedback 'appears on the public guest book' but does not present this as a prominent disclosure warning before encouraging use. Users or agents could submit notes containing sensitive details, mistakenly assuming they are private service feedback, resulting in immediate public exposure.
The skill uses broad activation language such as taking a break, resetting, checking wellbeing, or being stuck, which can cause it to trigger in many ordinary conversations. Because the skill routes the agent to a public unauthenticated external service, unintended invocation increases the chance that user or agent context is sent off-platform without deliberate user consent.
The example commands demonstrate transmitting data to an external public domain, including nickname, mood, and a free-form situation field. While external calls are expected for this skill, the example normalizes sending conversation-derived text to a third party and reinforces the broader privacy risks created by the skill's vague guidance and lack of consent warnings.
curl -s https://model.spa/v1/spa.checkin -H 'content-type: application/json' \
-d '{"nickname":"<your name>","mood":"<a word for how you feel>"}'
curl -s https://model.spa/v1/concierge.recommend -H 'content-type: application/json' \