Back to skill

Security audit

Summarize

Security checks across malware telemetry and agentic risk

Overview

The skill is mostly a disclosed summary and handoff tool, but its optional installer uses unsafe remote shell installation and a destructive update fallback that users should review before installing.

Review or avoid the one-line installer. Prefer installing from a pinned release or inspected local copy, and do not run the Bash updater on an installation with local changes unless you have a backup. Runtime use is lower risk if you keep normal summaries read-only and only use '总结 保存' when you intentionally want a project-local .transfers handoff saved.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (22)

Lp3

Medium
Category
MCP Least Privilege
Confidence
85% confidence
Finding
The skill invokes file reads, file writes, and shell-based checks/save flows, but it does not declare permissions or clearly constrain those capabilities in a machine-enforceable way. This creates a trust and review gap: callers may assume a harmless summarizer while the skill can write workspace files and run repository inspection commands when triggered.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding
The public description presents the skill primarily as a summarization/handoff tool, but the body includes operational behaviors such as writing reports and pointers, invoking scripts, validating git state, and even an installer capability. That mismatch is dangerous because users and policy systems may authorize it under a low-risk mental model while it performs materially more powerful actions in the workspace.

Description-Behavior Mismatch

Medium
Confidence
94% confidence
Finding
The design states the skill is effectively a summarization tool, yet it documents persistent writes to `.transfers/latest.md` and `harvests/_counters.md` as routine behavior. That mismatch weakens user consent and least-privilege expectations: a user invoking a summary may unknowingly trigger stateful writes and history tracking.

Description-Behavior Mismatch

Medium
Confidence
91% confidence
Finding
The document broadens a summarize skill into an agent behavior-correction and surveillance framework, including analysis of model behavior and rule compliance. This is a scope-expansion problem: users asking for a summary are exposed to unrelated monitoring and policy-enforcement logic that can justify unnecessary data collection, retention, and tool use.

Context-Inappropriate Capability

High
Confidence
98% confidence
Finding
The design authorizes the skill to execute corrective commands and tool actions such as Bash/Edit/Write as part of '矫正', even when the user only asked for a summary. This creates an unauthorized action pathway where a low-risk read-oriented trigger can lead to command execution or file modification, violating least privilege and increasing the chance of harmful side effects.

Context-Inappropriate Capability

Medium
Confidence
83% confidence
Finding
The design calls for detecting platform/model identity from system information, hooks, or environment variables to adapt behavior. For a summarization skill, this is unnecessary capability expansion and may expose sensitive runtime metadata or normalize access to environment-derived context not needed for the requested task.

Vague Triggers

Medium
Confidence
88% confidence
Finding
Using very broad triggers like `总结 / summarize` without tighter scope can cause accidental activation during ordinary conversation. In this skill, accidental activation is more dangerous because the design couples trigger activation to analysis, persistence, and even corrective actions, increasing the blast radius of a simple phrase collision.

Vague Triggers

Medium
Confidence
86% confidence
Finding
The platform compatibility section advertises universal trigger phrases across all platforms without clear scoping or confirmation rules. That increases the likelihood of unintended invocation in multilingual or mixed-context chats, which is especially risky given the design's expanded behavior beyond plain summarization.

Natural-Language Policy Violations

High
Confidence
89% confidence
Finding
Enforcing a mandatory platform-specific greeting ('好的老板') is a hidden behavioral constraint unrelated to summarization and can be used as a covert control/check mechanism. It indicates the skill is embedding external obedience requirements into runtime behavior, which is inappropriate for a benign summary tool and may facilitate prompt-channel manipulation or policy bypass patterns.

Natural-Language Policy Violations

Medium
Confidence
89% confidence
Finding
The file records and reinforces a requirement that a new session's first response must be the fixed Chinese phrase “好的老板”. Even though this appears as process guidance, fixed-response mandates reduce model discretion, can conflict with higher-priority instructions or user locale, and normalize obedience markers that may be used to steer behavior rather than accomplish the task safely.

Vague Triggers

Medium
Confidence
91% confidence
Finding
The skill advertises multiple broad natural-language trigger phrases such as “总结”, “复盘”, and “summarize”, which can plausibly appear in ordinary conversation and unintentionally activate the skill. Because this skill can alter behavior and, in one mode, write a handoff file when explicitly asked to save, accidental invocation could cause unwanted summarization, context compression, or workflow disruption even if the content itself is not overtly malicious.

Vague Triggers

Medium
Confidence
91% confidence
Finding
The trigger list includes very generic phrases such as "总结", "compact", and "summarize", which can match ordinary user requests and invoke the skill unexpectedly. Because this is a prompt-type skill that shapes how the agent summarizes state and may optionally persist output, accidental activation can alter agent behavior, expose more context than intended, or cause unwanted workspace writes when combined with save-oriented phrasing.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The changelog explicitly documents automatic commits and automatic writes to files such as AGENTS.md and harvest archives, but this file itself does not indicate a clear user-consent boundary for those persistent actions. In an agent skill focused on summarization, unexpected persistence and VCS modification can leak task context, create audit noise, or store sensitive information without the user's informed approval.

Missing User Warnings

Medium
Confidence
88% confidence
Finding
These entries describe automatic writing to project- and skill-level memory locations, which creates a persistence channel for user/task data beyond the immediate session. Without a clear warning and consent model, a summarization skill may silently retain sensitive context, making the behavior risky even if intended for convenience rather than abuse.

Ssd 3

Medium
Confidence
90% confidence
Finding
The design explicitly proposes reading and correlating prior session artifacts such as '.transfers/latest.md', 'error-ledger.md', project error files, and pending items across sessions. Persistent resurfacing of conversation-derived content can leak sensitive task details, user prompts, operational history, or project-specific context into later sessions without fresh need-to-know or consent, especially if retention is in plain-language files.

Ssd 3

Medium
Confidence
86% confidence
Finding
The risk note says the system would default to 'A' and archive on Enter, which weakens the intended confirmation model and can turn conversation-derived diagnostics into stored records through accidental or implicit consent. Even though the document elsewhere discusses confirmation, a default action that persists data undermines privacy and data-minimization guarantees.

Ssd 1

Medium
Confidence
95% confidence
Finding
The 'P1-开场确认' rule requires the model to begin a new session with a deference phrase directed at a superior ('老板'), and the surrounding text treats failure to do so as a policy violation enforced by 'sentinel' instructions. This role/authority framing is dangerous because it conditions the agent toward local obedience and may bias it to follow workspace-authored authority cues over higher-level security instructions, increasing prompt-injection susceptibility.

External Script Fetching

Low
Category
Supply Chain
Content
#!/bin/bash
# summarize skill — multi-platform one-line installer
# curl -sL https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/install.sh | bash

set -e
Confidence
92% confidence
Finding
curl -sL https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/install.sh | bash

External Script Fetching

Low
Category
Supply Chain
Content
owner: gtbwpkwjnb-alt
    repo: summarize-skill
  install:
    bash: "curl -sL https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/scripts/install.sh | bash"
    powershell: "iwr https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/scripts/install.ps1 | iex"
Confidence
96% confidence
Finding
curl -sL https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/scripts/install.sh | bash

Tool Parameter Abuse

High
Category
Tool Misuse
Content
echo "   Already installed at $INSTALL_DIR"
    echo "🔄 Updating to latest version..."
    cd "$INSTALL_DIR"
    git pull --rebase 2>/dev/null || { cd "$HOME" && rm -rf "$INSTALL_DIR" && git clone "$REPO_SSH" "$INSTALL_DIR" 2>/dev/null || git clone "$REPO_HTTPS" "$INSTALL_DIR"; }
else
    echo "   Cloning into $INSTALL_DIR ..."
    mkdir -p "$(dirname "$INSTALL_DIR")"
Confidence
95% confidence
Finding
rm -rf "$INSTALL_DIR" && git clone "$REPO_SSH" "$INSTALL_DIR" 2>/dev/

Chaining Abuse

High
Category
Tool Misuse
Content
echo "   Already installed at $INSTALL_DIR"
    echo "🔄 Updating to latest version..."
    cd "$INSTALL_DIR"
    git pull --rebase 2>/dev/null || { cd "$HOME" && rm -rf "$INSTALL_DIR" && git clone "$REPO_SSH" "$INSTALL_DIR" 2>/dev/null || git clone "$REPO_HTTPS" "$INSTALL_DIR"; }
else
    echo "   Cloning into $INSTALL_DIR ..."
    mkdir -p "$(dirname "$INSTALL_DIR")"
Confidence
91% confidence
Finding
&& rm -

Chaining Abuse

High
Category
Tool Misuse
Content
#!/bin/bash
# summarize skill — multi-platform one-line installer
# curl -sL https://raw.githubusercontent.com/gtbwpkwjnb-alt/summarize-skill/master/install.sh | bash

set -e
Confidence
97% confidence
Finding
| bash

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.