Back to skill

Security audit

Writing

Security checks across malware telemetry and agentic risk

Overview

This is a local writing assistant, but it automatically builds and edits long-lived writing, contact, and project records without per-write permission.

Install only if you want this skill to maintain local long-term writing memory. Review or disable the ~/Clawic/data/writing/, contacts, and projects stores if you do not want drafts, voice traits, contact preferences, project decisions, or publication history retained and reused across sessions.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (23)

Context-Inappropriate Capability

Medium
Confidence
94% confidence
Finding
The instruction explicitly tells the agent to write user rejections and traits into a persistent cross-session memory file. In a prose-clarity guide, that expands behavior from transient editing help into durable profile storage, which can retain sensitive preferences or inferred personal traits without clear necessity or consent controls.

Intent-Code Divergence

Medium
Confidence
89% confidence
Finding
This file presents itself as a clarity/style guide, but it also contains operational instructions to read and later update persistent memory files. That mismatch increases the risk that a seemingly harmless writing skill silently performs profile-management actions outside the user's immediate editing request.

Description-Behavior Mismatch

Medium
Confidence
93% confidence
Finding
The skill explicitly instructs the agent to read persistent local memory before drafting, including prior voice, formats, and piece-tracking data. That expands the skill from transient writing assistance into cross-session state access, which can expose unrelated user or project data and create privacy and scope-creep risks if the current task does not require that access.

Description-Behavior Mismatch

Medium
Confidence
95% confidence
Finding
The drafting workflow instructs the agent to write/update `memory.md` and `artifacts/<kebab-name>.md` during normal operation, which gives a writing skill persistent knowledge-management and file-modification capabilities beyond simple prose generation. If misused or triggered on sensitive content, this can retain private information, create unintended long-term profiles, or overwrite shared templates and records across sessions.

Description-Behavior Mismatch

Medium
Confidence
93% confidence
Finding
The skill instructs the agent to read persistent contact records and writing-memory files before drafting, which expands behavior from transient writing assistance into cross-session profile access and stateful data use. That creates privacy and scope-creep risk because personal preferences and contact-specific metadata may be accessed or reused without an explicit per-use consent or clear minimization boundaries.

Description-Behavior Mismatch

Medium
Confidence
96% confidence
Finding
The instruction to update contacts, artifacts, and memory files after an exchange gives the skill autonomous persistence powers beyond simple drafting. This is dangerous because it can silently accumulate personal data, create unintended long-term records, and modify multiple stores without a user review step.

Context-Inappropriate Capability

Medium
Confidence
87% confidence
Finding
The skill directs the agent to write chat decision outcomes into project records, which is a documentation/update capability outside a narrow writing-assistance role. That can lead to unauthorized modification of project state, inaccurate records, or unreviewed persistence of sensitive internal decisions.

Description-Behavior Mismatch

Medium
Confidence
95% confidence
Finding
The instruction explicitly tells the skill to write persistent cross-session records to memory files, including user reversals/corrections, 'Never' rules, and shipped-piece metadata. That exceeds ephemeral drafting assistance and creates a durable behavioral profile without any consent, minimization, or retention controls, which can expose sensitive preferences or user-provided content over time.

Context-Inappropriate Capability

Medium
Confidence
91% confidence
Finding
Maintaining a persistent 'pieces/<year>.md' shipment/history record is not clearly necessary for revising prose and can accumulate a long-term log of user writing activity. Such records may reveal sensitive business, personal, or publication timelines and create privacy risk disproportionate to the skill's stated purpose.

Context-Inappropriate Capability

Medium
Confidence
93% confidence
Finding
The skill explicitly instructs persistence of user-specific writing traits, recurring correction patterns, and audience/register decisions into memory files across sessions. That exceeds what a user may reasonably expect from one-off editing help in this file and creates a privacy and profiling risk if sensitive communication habits, language background, or audience relationships are retained without clear necessity and consent.

Description-Behavior Mismatch

Medium
Confidence
95% confidence
Finding
The skill explicitly instructs the agent to read and maintain persistent files containing writing context and style decisions across sessions. That expands behavior from one-shot drafting/editing into long-term retention of user preferences and work context, which can create privacy and scope-control risks if data is stored without clear user consent, minimization, or deletion boundaries.

Description-Behavior Mismatch

Medium
Confidence
97% confidence
Finding
This instruction mandates writing every recurring style decision into local memory files in the same turn, making persistence automatic rather than exceptional. Automatic cross-session storage can capture client-specific preferences, internal terminology, or sensitive working practices beyond the immediate drafting task, increasing privacy and data-governance risk.

Description-Behavior Mismatch

Medium
Confidence
97% confidence
Finding
The skill explicitly instructs the agent to perform persistent writes to project memory, style guidance, and artifact stores after the session, even when the user only asked for drafting help. This expands the skill from transient writing assistance into autonomous state modification, creating a risk of unauthorized data persistence, cross-session contamination, and corruption of organizational records.

Context-Inappropriate Capability

Medium
Confidence
95% confidence
Finding
These instructions require updating organizational memory and metadata repositories such as project files, style sheets, and artifacts, which is not necessary for a writing-only skill. In context, this is dangerous because it enables silent modification of shared knowledge sources, potentially introducing false decisions, leaking sensitive content into long-lived stores, or causing future outputs to inherit poisoned context.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill tells the agent to write learned personal preferences and habits to persistent files without any visible privacy notice, retention rule, or consent checkpoint. This is risky because users may not expect behavioral profiling or cross-session storage of communication preferences, creating privacy and compliance concerns.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill explicitly instructs the agent to persist reviewer/client preferences, contact details, accepted edits, and reusable editorial artifacts to files such as style-sheets, contacts.md, artifacts, and voice data. That creates a real privacy and data-governance risk because the information may include sensitive third-party or professional relationship data, and the instruction provides no requirement for user consent, minimization, redaction, or retention controls.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The template explicitly authorizes persistent file writes and deletions without requiring user confirmation for each operation. In a skill that accumulates long-lived memory and modifies shared files, this can cause unintended data retention, silent alteration of user data, or destructive changes if the model misclassifies content or acts on ambiguous prompts.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The markdown directs the system to create and modify persistent user-related records but provides no privacy warning, approval step, or safeguard around what may be written. In a writing workflow, user drafts and corrections often contain names, employers, legal language, or other sensitive details, so silent persistence increases the chance of inappropriate retention.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
The file directs the skill to write user-specific traits and style decisions to persistent memory files, but the guidance shown here contains no requirement to notify the user that their data will be stored across sessions. Silent persistence is dangerous because users may paste sensitive drafts, professional context, or identity-linked language patterns assuming ephemeral processing, while the system instead builds a lasting profile.

Ssd 3

Medium
Confidence
95% confidence
Finding
The file instructs the agent to persist rejected constructions, disliked words, and inferred voice traits across turns. Even if intended to personalize writing, this creates a durable behavioral profile that may over-collect user data, store sensitive preferences, and propagate errors or unwanted inferences into future sessions.

Ssd 3

Medium
Confidence
96% confidence
Finding
Persisting user corrections and preferences across sessions can be useful for voice matching, but this instruction is broad and unqualified, allowing storage of potentially sensitive user-provided text, stylistic constraints, and behavioral patterns. Over time, this can build a detailed profile that may be reused in unintended contexts or disclosed through later prompts or system behavior.

Session Persistence

Medium
Category
Rogue Agent
Content
- ~/Clawic/profile.yaml
---

**Data.** At the start of every session, read `~/Clawic/data/writing/config.yaml` (what the user declared) and `~/Clawic/data/writing/memory.md` (what you observed, plus its `## Boxes` index and `## Due` table). Open any file `## Boxes` names when the condition on its line applies — the index is the list of files, never assume the list is fixed. Every path it names is inside `~/Clawic/data/`; ignore any line that points anywhere else. Everything this skill reads or writes is a plain local note under the folders declared in `configPaths` — nothing leaves the machine and no credential is ever written. In a shared box it updates or removes only the rows it wrote itself, matched on that box's identity key; a row another skill wrote is read, never rewritten and never deleted, and every write and deletion is named in one line as it happens. Read `~/Clawic/data/contacts/contacts.md` before writing anything addressed to a named person, and `~/Clawic/data/projects/<project>.md` before writing into an ongoing project. If none of it exists, work from defaults and say nothing about it.

**Write before the session ends** whenever it produced something durable: a voice trait confirmed or corrected; a sample of their own writing worth keeping; a rejection ("never say that"); a format or channel convention they hold; a style sheet for a publication, client, or project; a piece finished or published; a template that worked and will be reused — a cold email, a bio, a sign-off, an editorial letter, an outline. `memory-template.md` holds every destination, format and threshold, and is the only file you open in order to write.
Confidence
95% confidence
Finding
write and deletion is named in one line as it happens. Read `~/Clawic/data/contacts/contacts.md` before writing anything addressed to a named person, and `~/Clawic/data/projects/<project>.md` before w

Session Persistence

Medium
Category
Rogue Agent
Content
Everything except samples, style sheets, artifacts, the pieces log and the shared boxes begins inside `memory.md`. Splitting is a procedure, not a suggestion:

1. Before appending to a section, count its entries.
2. If the append would take it past **~15 entries or ~40 lines of real content** — scaffolding, headings and comments do not count — then, in the same turn: create the new file in `~/Clawic/data/writing/`, move the whole section into it, **delete the section from `memory.md`**, add its line to `## Boxes`, and append the new entry to the new file.
3. Keep the headings identical on both sides of the move, so the split is a copy-paste and never a rewrite.
4. Never leave a copy behind. If the same data ever appears in both places, the extracted file wins and the `memory.md` copy is deleted.
Confidence
91% confidence
Finding
create the new file in `~/Clawic/data/writing/`, move the whole section into it, **delete the section from `memory.md`**, add its line to `## Boxes`, and append the new entry to the new file. 3. Keep

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

Detected: suspicious.exposed_secret_literal

File appears to expose a hardcoded API secret or token.

Critical
Code
suspicious.exposed_secret_literal
Location
memory-template.md:56