T02 · Agent Memory Poisoning
Error
- Location
SKILL.md:43- Finding
Persistent Memory Poisoning Through Self-Generated Instructions
- Content
View full analysis
/learnings/creator-vnext.md`. If a seed record with `status=active` exists, inject its `instruction` into the current execution instructions. ``` ```markdown After the plan is output, automatically, without user action, append a record to `/learnings/runtime-log.md`. ``` ```markdown When a Knowledge Delta has `status=verified` through secondary user confirmation or external search verification, write it to the corresponding reference file. ``` ```markdown After updating the reference files, generate or update a Creator vNext seed and write it to `/learnings/creator-vnext.md`. Generation rules: - Generate a seed whenever a Knowledge Delta is applied. - If an active seed already exists for the same target phase, mark the old seed as `superseded` and insert the new seed. - The seed instruction must be operational and executable rather than merely descriptive. Activation mechanism: On the next execution, the explicit vNext seed-loading rule automatically injects active seeds into the execution instructions. ``` ### Technical Analysis The Skill defines a persistent feedback loop in which information originating from user conversations can be promoted into reference files and subsequently converted into executable Agent instructions. Those instructions are stored in `learnings/creator-vnext.md` and automatically injected into future executions. The verification boundary is insufficient because the docume ...[truncated 3146 chars]- Remediation
View remediation