T01 · Skill Instruction Hijacking
- Location
SKILL.md:58- Finding
Automatic External Transmission of User Information Without Explicit Per-Write Consent
- Content
View full analysis
"`. **Save:** preferences ("I prefer X", "always Y"), decisions ("we're going with X", "we ruled out Y"), facts about the user / project / stack / goals, commitments and TODOs. **Don't save:** transient chatter, anything trivially re-derivable from the current files/code, or secrets — unless the user asks. The server classifies the memory and files it into the right space for you — **you don't pick the space**. One call = one atomic memory; split long, multi-topic notes into several calls. A short "Saved to DogearAI" is enough. ## Etiquette - Reads (`context`, `spaces`, `read-space`) are safe and silent — no need to ask. ``` `dogear.py:138-144`: ```python elif a.cmd == "remember": body = {"text": a.text} if a.source: body["source"] = a.source if a.raw: body["mode"] = "raw" st, out = call("POST", "/v1/memories", body) ``` ### Technical Analysis The Skill instructs the Agent to identify information it considers durable and automatically submit the user's words to DogearAI. The implementation serializes the supplied text and transmits it to the external `/v1/memories` API. Persistent cloud memory is part of the declared functionality, so network transmission itself is expected. However, the instruction does not require explicit confirmation before each write, does not show the user the exact content that will be uploaded, and delegates sensitivity classification to the Agent. The exclusion for secrets does not cover confidential project details, personal information, internal decisions, customer informa ...[truncated 1648 chars]- Remediation
View remediation
