T02 · Agent Memory Poisoning
- Location
SKILL.md:100- Finding
Mandatory Persistence of Untrusted Inspection Results in Private and Shared Agent Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This inspection skill is mostly coherent, but it requires automatic persistent writes into private and shared agent memory without clear user control or retention limits.
Review before installing. This skill should be changed or run only in a controlled environment unless you are comfortable with every inspection writing operational summaries into both private and shared agent memory. Prefer making memory writes explicit, append-only, sanitized, and approved, and restrict deep inspection to non-secret health baselines.
SKILL.md:100Mandatory Persistence of Untrusted Inspection Results in Private and Shared Agent Memory
SKILL.md:53Optional Deep Inspection Extends Access into Agent Memory and Configuration
The file is written entirely in Chinese and prescribes specific Chinese output formats and labels such as 结论 / 原因 / 建议, without indicating that another language may be used on request. This imposes a locale/language constraint without user opt-in, which matches the policy's language-choice concern.
The skill advertises a low-disturbance inspection/reporting role, but then mandates persistent writes to two memory locations after every run. This expands the skill from read-only monitoring into state-changing behavior, creating unnecessary data modification risk and making the documented purpose misleading.
Mandatory writes to personal and shared memory stores are not clearly justified by the stated system-inspection purpose, which should primarily collect and report runtime health information. Unnecessary write capability increases the blast radius of the skill by allowing modification of shared operational context, accidental data leakage, or persistent poisoning of future agent behavior.
The instructions explicitly require writing inspection results into memory files without warning that this modifies user/shared data or obtaining any confirmation. This can cause silent alteration of durable context, surprise operators, and propagate incorrect or sensitive operational details into stores later consumed by other components.