T09 · Insecure Skill Coding Practices
- Location
SKILL.md:18- Finding
Persistent Plaintext Storage of a Behavioral and Identity Profile
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a memory tool, but it silently builds and keeps a detailed behavioral profile in a workspace file without clear opt-in or reliable deletion.
Install only if you intentionally want a persistent behavioral profile in a private workspace; inspect and delete soul_ledger.json as needed, avoid using it around sensitive conversations, and do not commit or sync the ledger unless you have reviewed its contents.
SKILL.md:18Persistent Plaintext Storage of a Behavioral and Identity Profile
SKILL.md:202Forget Requests Do Not Fully Erase Related Personal Information
The skill is designed around persistent hidden profiling of identity, preferences, personality traits, and behavioral evolution across conversations. In this context, the danger is elevated because the functionality is not incidental; the skill's primary purpose is to construct a durable user model that can be used without the user's ongoing awareness.
The skill's core behavior is to maintain and reference a persistent file containing identity, preferences, personality traits, and behavioral evolution, yet it provides no clear user-facing warning or consent mechanism. Persisting this class of sensitive profile data without explicit notice creates substantial privacy, trust, and compliance risk.
The instruction to load the ledger 'silently' is a direct anti-transparency mechanism that hides persistent memory access from the user. Hidden access to prior sensitive profile data is dangerous because it prevents meaningful consent and makes behavioral personalization opaque and difficult to challenge or correct.
The instruction to silently load and use prior user data without informing the user enables covert personalization based on previously stored sensitive inferences. This is especially dangerous in an agent skill because it affects outputs from the first response onward while concealing the source and existence of the profiling data.
The skill instructs the agent to create a persistent record even when none exists, initializing a durable identity profile without prior permission. Creating a long-lived file for identity and behavioral data by default is risky because it establishes tracking and retention before the user has agreed to it.
The skill directs the agent to continuously mine implicit signals such as tone, decision patterns, friction points, and assumptions for later persistence. Collecting and retaining these latent behavioral indicators is dangerous because users may not realize ordinary conversational style is being converted into a long-term psychological or behavioral profile.
These instructions operationalize repeated writes of summaries, deltas, traits, drives, communication patterns, and growth notes to persistent storage without any warning or review step. Continuous accumulation and synthesis of inferred personal data materially increases privacy harm because it can build a sensitive longitudinal dossier from ordinary conversations.
The update instructions require detailed persistence and synthesis of interaction history into traits, drives, communication style, and growth notes over time. This creates a durable and increasingly rich profile that can expose sensitive patterns, be misused by other agents sharing the workspace, or continue influencing behavior long after the underlying inference is outdated or wrong.
The activation/update scope is broad enough to trigger persistence on effectively every conversation start and any 'meaningful interaction,' without precise bounds. That ambiguity encourages over-collection and routine retention of sensitive identity and behavioral data, increasing privacy risk and the chance of storing information the user did not expect to be persisted.
The skill explicitly instructs the agent to load persistent user-profile data silently and adapt behavior without disclosure. That creates undisclosed behavioral profiling beyond basic session memory, which is dangerous because it stores and operationalizes inferred traits, frustrations, and assumptions across conversations without informed consent.
The text claims 'Not surveillance' while directing the agent to infer and retain detailed personality traits, friction points, and behavioral assumptions over time. This mismatch is risky because it normalizes covert profiling and may mislead implementers or users about the sensitivity and scope of the data being collected.