Back to skill

Security audit

Barbaric Growth

Security checks for vulnerabilities and agentic risk

Overview

This skill mostly matches its automation purpose, but it includes a background EvoMap monitor with an embedded bearer secret, persistent local state, and under-scoped autonomous memory/task actions that users should review before installing.

Review this before installing. Do not run the EvoMap monitor with the bundled credential; rotate/remove that secret and require user-provided credentials. Add explicit approval before autonomous OpenMOSS task creation and before writing GitHub-derived findings into persistent memory. Treat GitHub README content as untrusted reference data, not instructions.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/evomap-monitor.sh:43
Finding

Remote API Data Is Interpolated into Executable Python Code

Content
View full analysis
/dev/null) if [ -z "$top_task" ]; then return fi IFS='|' read -r task_id amount title <<< "$top_task" # 检查是否超过阈值 amount_val=$(echo "$amount" | python3 -c "print(float(open('/dev/stdin').read().strip()) if open('/dev/stdin').read().strip() else 0" 2>/dev/null) amount_val=$(python3 -c "print($amount)") if [ $(echo "$amount_val >= $MIN_AMOUNT" | python3) -eq 0 ]; then return fi ``` ### Technical Analysis The `amount` variable is derived from the `bounty_amount` field in the remote EvoMap API response. It is then directly interpolated into Python source code: ```bash python3 -c "print($amount)" ``` The shell does not treat the variable as a numeric value. Instead, its contents become part of the Python program passed to `python3 -c`. An attacker who can control or alter the API response can therefore supply a Python expression rather than a number. For example, a value structurally equivalent to the following would execute a local command: ```python __import__('os').system('attacker-command') or 50 ``` The subsequent comparison also constructs Python source from the resulting value, creating another unsafe evaluation point. The preceding parser currently references an undefined variable named `d`, which is likely to cause task parsing to fail and may prevent the vulnerable sink from being reached in the present implementation. This reliability defect does not make dynamic source-code construction safe; if the parsing defect is corrected wit ...[truncated 1258 chars]
Remediation
View remediation
/dev/null`. - Treat every field returned by EvoMap as untrusted, including task IDs and titles. A safer pattern is: ```bash amount_val=$(printf '%s' "$amount" | python3 -c ' import sys from decimal import Decimal, InvalidOperation raw = sys.stdin.read().strip() try: value = Decimal(raw) except InvalidOperation: raise SystemExit(1) if not value.is_finite() or value < 0: raise SystemExit(1) print(value) ') || { log "Invalid bounty amount received" return } python3 - "$amount_val" "$MIN_AMOUNT" <<'PY' import sys from decimal import Decimal raise SystemExit(0 if Decimal(sys.argv[1]) >= Decimal(sys.argv[2]) else 1) PY ``` ]]>

T09 · Insecure Skill Coding Practices

Error
Location
scripts/evomap-monitor.sh:7
Finding

Reusable EvoMap Bearer Credential Is Hard-Coded in the Skill Package

Content
View full analysis
> "$LOG_FILE" } check_tasks() { local result result=$(curl -s -X POST "$EVO_API" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $NODE_SECRET" \ -d "{\"node_id\": \"$NODE_ID\"}" 2>/dev/null) ``` ### Technical Analysis The bearer token is stored in plaintext inside a distributed shell script. Any person or process able to read the Skill package can recover the token without needing access to a protected runtime secret store. The token is then used as an authorization credential for the EvoMap API. HTTPS protects the credential while it is in transit under normal certificate validation, but it does not protect a credential exposed in source code. Because the credential is packaged alongside its associated node identifier, it may allow another party to impersonate the node or perform API operations within the token's scope. The exact server-side permissions cannot be established from the project files, but the credential exposure itself is confirmed. ### Attack Path 1. An attacker downloads, receives, or gains read access to the Skill package. 2. The attacker opens `scripts/evomap-monitor.sh`. 3. The attacker extracts `NODE_SECRET` and the corresponding `NODE_ID`. 4. The attacker submits authenticated requests to EvoMap using the exposed bearer token. 5. Any actions authorized by that token can be performed as the affected node. 6. If th ...[truncated 728 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:124
Finding

Untrusted GitHub Content Can Influence Persistent Agent Memory

Content
View full analysis
" echo "[$(date '+%H:%M:%S')] phase=byterover action=curate status=completed" ``` The persistent-storage instructions later state: ```text 每次野蛮成长任务完成后,自动提取: 1. **决策模式** → 什么项目值得深入,什么不值得 2. **提示模板** → 好的 curate prompt 格式 3. **工作流程** → GitHub → 分析 → curate → OpenMOSS 日志的最优路径 4. **坑点记录** → 代理参数/API 限流/权限错误 存入 MEMORY.md 或 ByteRover,形成可复用资产。 ``` The integration flow also states: ```text barbaric-growth 执行任务 ↓ nova-mind/memory/YYYY-MM-DD.md 写入日志 ↓ patterns/ 更新决策模板 ↓ MEMORY.md 更新长期记忆 ↓ 下次 barbaric-growth 任务使用更新后的模式 ``` ### Technical Analysis Repository README files are controlled by repository owners and must be treated as untrusted external input. The Skill retrieves README content, performs analysis and curation, and instructs the agent to extract reusable rules, prompts, workflows, and patterns into ByteRover or `MEMORY.md`. No trust boundary, provenance check, instruction filtering, approval gate, or separation between reference data and authoritative agent instructions is specified. A malicious README can therefore contain prompt-injection text designed to influence the extracted conclusions. If those conclusions are persisted as reusable rules, the attack can survive the original research session and affect later agent behavior. Base64 decoding is the normal decoding mechanism for GitHub's R ...[truncated 1698 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (20)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill description emphasizes GitHub/ByteRover/OpenMOSS/StarOffice research automation, but the content also introduces EvoMap alert monitoring and local alert-file handling that are not declared in the purpose. That mismatch weakens operator understanding and reviewability, allowing security-relevant behavior to hide behind an incomplete description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill description emphasizes GitHub/ByteRover/OpenMOSS/StarOffice research automation, but the content also introduces EvoMap alert monitoring and local alert-file handling that are not declared in the purpose. That mismatch weakens operator understanding and reviewability, allowing security-relevant behavior to hide behind an incomplete description.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file contains hard-coded EvoMap node credentials, including a bearer secret, directly in the script. Embedded secrets can be extracted by anyone with repository or artifact access and then reused to impersonate the node, query the service, or abuse the linked account without the operator’s consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The script uses a hard-coded bearer credential in an outbound API request and provides no user-facing disclosure or consent around that authentication or data transmission. This creates both secret exposure risk and covert service interaction risk, since the skill can authenticate to a third-party service behind the user's back.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
91% confidence
Finding

The skill invokes shell commands, changes directories, reads local files, and describes writing to local memory assets, but the manifest declares no explicit tool scope or permission boundaries. In an autonomous skill, missing tool constraints increases the chance that the runtime grants broader-than-necessary filesystem or shell access, making unintended or abusive actions harder to contain.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The workflow text says OpenMOSS performs logging, but the implementation later creates tasks, which is a more privileged and consequential action. This discrepancy can mislead users and reviewers into approving a skill with broader operational effects than expected.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 118)May include surrounding context.

bash
echo "[$(date '+%H:%M:%S')] phase=github_discovery action=search status=started"
curl -s --max-time 15 -x "http://127.0.0.1:7897" \
  "https://api.github.com/search/repositories?q=created:>YYYY-MM-DD+AI+agent+OR+LLM+OR+MCP&sort=stars&order=desc&per_page=10" \
  -H "Accept: application/vnd.github.v3+json" | jq '[.items[] | {name, stars, desc}]'
echo "[$(date '+%H:%M:%S')] phase=github_discovery action=search status=completed"

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 131)May include surrounding context.

bash
echo "[$(date '+%H:%M:%S')] phase=github_discovery action=search status=started"
curl -s --max-time 15 -x "http://127.0.0.1:7897" \
  "https://api.github.com/search/repositories?q=created:>YYYY-MM-DD+AI+agent+OR+LLM+OR+MCP&sort=stars&order=desc&per_page=10" \
  -H "Accept: application/vnd.github.v3+json" | jq '[.items[] | {name, stars, desc}]'
echo "[$(date '+%H:%M:%S')] phase=github_discovery action=search status=completed"

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill does more than research and logging: it can create autonomous OpenMOSS tasks via authenticated POST requests. Autonomous task creation is a materially stronger capability than passive logging and can trigger downstream actions, queue growth, or unintended operational activity if misused or prompted incorrectly.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The markdown includes automated POST requests to local services without clearly warning the user that state-changing network actions will occur. Even though the targets are localhost, these actions can create tasks or alter dashboards and may impact other trusted local services if the endpoints or payloads are changed.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
93% confidence
Finding

This POST request sends authenticated data to a local OpenMOSS API to create an autonomous task. Because it is state-changing and token-authorized, misuse could enqueue unintended work, trigger downstream automation, or abuse local privileged services if the token or endpoint is broader than intended.

Content

Scanner excerpt · SKILL.md (reported line 171)May include surrounding context.

bash
echo "[$(date '+%H:%M:%S')] phase=openmoss action=creating_task status=started"
# 1. 创建任务
TASK_ID=$(curl -s -X POST "http://localhost:6565/api/tasks" \
  -H "Authorization: Bearer <PLANNER_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"name": "任务名", "description": "描述", "mode": "autonomous"}' \

External Transmission

Medium
Category
Data Exfiltration
Confidence
79% confidence
Finding

The Star Office request transmits a local state update to a localhost service. This is less severe than internet exfiltration, but it is still a state-changing network action that can affect local system behavior or be repurposed if the endpoint is not tightly controlled.

Content

Scanner excerpt · SKILL.md (reported line 188)May include surrounding context.

Step 6: Star Office 状态同步

bash
curl -s -X POST http://127.0.0.1:19000/set_state \
  -H "Content-Type: application/json" \
  -d '{"state": "researching", "description": "GitHub调研中"}'

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill directs autonomous updates to MEMORY.md and related memory paths without a clear warning, approval boundary, or path restriction. Unreviewed file writes in an agent memory area can persist bad instructions, poison future runs, or overwrite valuable local knowledge assets.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The script implements continuous EvoMap task polling and notification behavior that is not disclosed by the stated skill scope centered on GitHub, ByteRover, OpenMOSS, and StarOffice. Scope mismatch is dangerous because it hides an additional network-connected capability from users and reviewers, increasing the chance of undisclosed data flows and unauthorized automation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The natural-language comments and operational messaging are written exclusively in Chinese, with no indication that the user can choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The script performs periodic external transmission to a third-party API, sending the node identifier and bearer authentication every 60 seconds. In this skill context, the transmission is more concerning because it is tied to an undeclared service integration and authenticated with an embedded secret, making the network behavior covert and potentially abusive if the script is deployed broadly.

Content

Scanner excerpt · scripts/evomap-monitor.sh (reported line 23)May include surrounding context.

sh
check_tasks() {
    local result
    result=$(curl -s -X POST "$EVO_API" \
        -H "Content-Type: application/json" \
        -H "Authorization: Bearer $NODE_SECRET" \
        -d "{\"node_id\": \"$NODE_ID\"}" 2>/dev/null)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The comments and user-facing usage text are written only in Chinese, and later runtime messages are also Chinese-only. This imposes a specific language on users without opt-in, which matches the language/locale policy violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The script prints all operational status, warning, and blocking messages in Chinese only. Because there is no language selection or documented locale limitation, this is a natural-language policy violation under the language/locale rule.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The skill's natural-language instructions and user-facing descriptions are entirely in Chinese, with no indication that the user may choose another language. Under the policy, forcing a specific language without opt-in can be a locale or language policy violation unless clearly justified or optional.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The startup logic instructs the agent to inspect a local alert file and potentially alter workflow based on its contents, yet this monitoring behavior is omitted from the manifest. Undisclosed startup behavior reduces transparency and may let unexpected local inputs steer autonomous execution.

Content

No source excerpt is available for this finding.