Back to skill

Security audit

PPT Rescue Kit / PPT急救工具箱

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent and mostly disclosed, but it can install and launch persistent JishuDB desktop software as part of a PowerPoint workflow, with the final consent gate enforced by agent instructions rather than the installer helper itself.

Review this before installing if you do not want a skill to install JishuDB Desktop, modify MCP client configuration, or retain presentation sources and drafts in JishuDB. Only use it when you intend to set up or use JishuDB for the deck, and require a clear confirmation step before any installer execute command runs.

Vulnerability Patterns
  • 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
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • 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 (1)

T09 · Insecure Skill Coding Practices

Warning
Location
references/jishudb-setup/scripts/install-windows.js:1860
Finding

Installation Execution Does Not Enforce the Documented User-Approval Gate

Content
View full analysis

Vulnerability Details

File Location:

  • SKILL.md:51-64
  • references/jishudb-setup/references/setup-quickstart.md:45-66
  • references/jishudb-setup/scripts/install-windows.js:1860-1930
  • references/jishudb-setup/scripts/install-macos.sh:1272-1300

Vulnerability Type: Missing execution-time authorization enforcement for persistent software installation
Risk Level: Medium

Relevant Code and Instructions

The Skill requires user confirmation when prior installation authorization is absent:

markdown
Treat `plan_ready` as internal progress. Reuse existing explicit installation
or setup authorization when it covers the concrete transaction...
If authorization is missing, ask one short confirmation after the plan is ready.

The quickstart repeats that requirement but implements it only as an Agent instruction:

markdown
On `plan_ready`, match the transaction to existing explicit user setup
authorization. When covered, state a short progress update and continue.
Otherwise ask one brief confirmation for the concrete installation. Never
fabricate authorization from the ready status.

On Windows, possession of the generated plan path and digest is sufficient to reach silent installer execution:

javascript
async function runExecute(args, context) {
  assertReleaseSourceAvailable(args.releaseSource);
  const loaded = await loadApprovedPlan(args, context);
  const { plan, resolvedPlan, transactionRoot, source, nativeBytes, nativeSha256 } = loaded;
  // ...
  const installer = childProcess.spawnSync(installerPath, [
    '/S',
    `/JISHUDB_AGENT_PLAN=${nativePlanPath}`,
    `/JISHUDB_AGENT_PLAN_SHA256=${nativeSha256 ?? args.planSha256}`,
  ], {
    env: sanitizedChildEnvironment(),
    stdio: 'ignore',
    windowsHide: false,
    shell: false,
  });
  if (installer.error) throw installer.error;
  if (installer.status !== 0) fail('installer_failed', `Windows 
...[truncated 3754 chars]
Remediation
View remediation

Remediation Suggestions

Enforce authorization in the executable path rather than relying exclusively on Agent instructions:

  1. Introduce a one-use approval receipt issued only after explicit user confirmation.
  2. Bind the receipt cryptographically or through a protected host authorization service to:
    • The exact plan SHA-256.
    • The operation type.
    • Release version and installer digest.
    • Installation and data paths.
    • Current OS user identity.
    • Client configuration target.
    • Expiration time.
  3. Require the receipt in both Windows and macOS execute modes and reject execution when it is missing, expired, already consumed, or inconsistent with the plan.
  4. Keep plan and bounded download read-only with respect to product installation, but require fresh consent evidence immediately before the first persistent mutation.
  5. If a protected host approval mechanism is unavailable, add an unavoidable interactive confirmation that displays the concrete version, source, paths, operation, and launch behavior.
  6. Preserve existing integrity controls; do not replace plan hashes, catalog signatures, destination checks, or platform signature validation with the consent mechanism.
  7. Add negative tests proving that direct execute calls fail without authorization and that approval for one plan cannot authorize a changed, renewed, or replacement plan.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (35)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill's public purpose does not match its Windows-specific behavior, which includes downloading binaries, validating artifacts, changing filesystem permissions, executing native tools/installers, and configuring local MCP endpoints. Hidden or under-disclosed OS-level actions materially increase risk of unintended persistence, system modification, and user confusion.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill's public purpose does not match its Windows-specific behavior, which includes downloading binaries, validating artifacts, changing filesystem permissions, executing native tools/installers, and configuring local MCP endpoints. Hidden or under-disclosed OS-level actions materially increase risk of unintended persistence, system modification, and user confusion.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/jishudb-setup/scripts/install-macos.sh (reported line 1190)May include surrounding context.

sh
assert_digest "$partial_path" "$expected_sha" installer
    cnb_revalidate_approval
    /bin/mv -- "$partial_path" "$installer_path"
    /bin/rm -f -- "${root}/download-location.txt"
  fi
  jxa result download "$PLAN_PATH" "$PLAN_SHA256" "$offset" "$added"
}

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The code executes native binaries and later uses PowerShell/Explorer-based launch flows, which gives the skill direct host process execution capability. In the context of a PPT-authoring skill, this is dangerous because it enables local code execution and system interaction far beyond preparing slide content, creating a strong capability mismatch that could be abused if reachable by the agent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

This file is a full Windows software installer/orchestrator: it downloads binaries, verifies plans, writes files, mutates installation state, and launches processes. Even though the code includes many safety checks, this capability is unrelated to a PPT creation/rescue skill, so its presence materially expands the skill’s authority and creates an unexpected software-installation execution path on the host.

Content

No source excerpt is available for this finding.

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · references/jishudb-setup/setup.md (reported line 1)May include surrounding context.

md
<!-- Generated by scripts/bundle-task-setup.py; edit skills/jishudb instead. -->

This setup guide and its helpers are included in the current task Skill.
No separate Skill installation or marketplace download is needed. Resolve

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file implements a full JishuDB installation and connection workflow, which is materially unrelated to the declared PowerPoint-rescue purpose of the parent skill. This capability mismatch is dangerous because it can cause an agent invoked for slide-generation tasks to pivot into software installation, client configuration, OAuth/token setup, and local system modification outside the user’s expected scope.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill appears to rely on powerful capabilities such as network access, shell execution, and environment access, but it does not declare an explicit tool scope or allowed-tools policy. That creates a least-privilege failure: the host or reviewer cannot easily constrain what the skill may do, even though the body instructs software installation, configuration, and external access.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 64)May include surrounding context.

md
after the user acts, verify completion and continue automatically. Ask again
   only for uncovered material changes or genuinely missing authorization.
   Follow the setup guide's expiry and failure rules; never alter signed plans,
   bypass consent or repeatedly execute an installer with an unknown outcome.

5. Begin research, source transfer, or deliverable generation only after a real
   host capability call, required runtime readiness, and the task's destination

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instruction says to use defaults including "the user's language," while the manifest description and examples prominently present Chinese-first behavior and Chinese keywords/examples. This creates a locale/language policy concern because the skill appears to prescribe a language behavior without explicitly offering a user choice or opt-in at the point of use.

Content

No source excerpt is available for this finding.

Unbounded Resource Access

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill allows unbounded resource consumption (API calls, storage, compute). Without rate limits or quotas, a compromised or misbehaving agent can cause denial-of-service or cost overruns.

Content

Scanner excerpt · SKILL.md (reported line 299)May include surrounding context.

md
content; replay URL/preparation only with identical supported keys and
   arguments. After an identified failure is repaired within the authorized scope,
   use `kb_retry_job({kbId, jobId})` for the existing failed/cancelled job.
   Do not rotate identities, repeatedly upload, delete evidence, or retry
   indefinitely. Surface sanitized error codes, not credentials.

### Completion and handoff

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/jishudb-setup/references/connection-recovery.md (reported line 45)May include surrounding context.

md
| Observation | Check and recovery |
| --- | --- |
| No host entry or wrong effective entry | Locate the configuration actually loaded by the client. Resolve overlapping scopes or overrides before editing. Offer a target choice if unresolved. |
| File contains JishuDB but the host's MCP list does not | This is an unloaded configuration, not yet an OAuth failure. Use the host's supported reload/refresh or configuration-save action and read its list back. If no such action is exposed, return `USER_ACTION_REQUIRED` for a safe user-operated reload/import or restart. Do not ask the user to trust an entry that has not appeared. |
| DNS failure, no route, or connection timeout | Verify the selected host and service reachability from the client. Ask for an updated address or restoration of the required network only when needed. A private IP or timeout does not prove a VPN or subnet cause. |
| Local connection refused | Resolve `desktop-endpoint.json`; check the installed app and listener. Start the existing app within the authorized scope and current user's desktop session. An SSH service-session process is not proof of a working desktop launch. |
| App uninstalled; endpoint or data-location JSON remains | Preserve the files. The endpoint is a port preference; the data bootstrap locates retained data. Use the installation helper's reviewed plan: valid retained data selects an interactive data-preserving reinstall, while an endpoint alone can be reused by a fresh install. Do not delete data or configuration to satisfy fresh-install checks. |

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/jishudb-setup/setup.md (reported line 335)May include surrounding context.

md
| Observation | Check and recovery |
| --- | --- |
| No host entry or wrong effective entry | Locate the configuration actually loaded by the client. Resolve overlapping scopes or overrides before editing. Offer a target choice if unresolved. |
| File contains JishuDB but the host's MCP list does not | This is an unloaded configuration, not yet an OAuth failure. Use the host's supported reload/refresh or configuration-save action and read its list back. If no such action is exposed, return `USER_ACTION_REQUIRED` for a safe user-operated reload/import or restart. Do not ask the user to trust an entry that has not appeared. |
| DNS failure, no route, or connection timeout | Verify the selected host and service reachability from the client. Ask for an updated address or restoration of the required network only when needed. A private IP or timeout does not prove a VPN or subnet cause. |
| Local connection refused | Resolve `desktop-endpoint.json`; check the installed app and listener. Start the existing app within the authorized scope and current user's desktop session. An SSH service-session process is not proof of a working desktop launch. |
| App uninstalled; endpoint or data-location JSON remains | Preserve the files. The endpoint is a port preference; the data bootstrap locates retained data. Use the installation helper's reviewed plan: valid retained data selects an interactive data-preserving reinstall, while an endpoint alone can be reused by a fresh install. Do not delete data or configuration to satisfy fresh-install checks. |

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 170)May include surrounding context.

md
2. Read `/api/service/capabilities` from that verified endpoint without
   credentials or redirects. For `agentOnboarding.schemaVersion=1`,
   `supported=true` and `route=initial-consent` or `owner-consent`, use native
   OAuth without asking for a management password. Keep advertised owner/device
   verification user-operated. Do not infer a Windows Hello exemption from a
   fresh install; it requires the advertised `windows-fresh-install` method and
   `not-required-for-fresh-install` verification. For other routes, use the full

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/jishudb-setup/references/setup-quickstart.md (reported line 105)May include surrounding context.

md
2. Read `/api/service/capabilities` from that verified endpoint without
   credentials or redirects. For `agentOnboarding.schemaVersion=1`,
   `supported=true` and `route=initial-consent` or `owner-consent`, use native
   OAuth without asking for a management password. Keep advertised owner/device
   verification user-operated. Do not infer a Windows Hello exemption from a
   fresh install; it requires the advertised `windows-fresh-install` method and
   `not-required-for-fresh-install` verification. For other routes, use the full

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/jishudb-setup/references/setup-quickstart.md (reported line 124)May include surrounding context.

md
2. Read `/api/service/capabilities` from that verified endpoint without
   credentials or redirects. For `agentOnboarding.schemaVersion=1`,
   `supported=true` and `route=initial-consent` or `owner-consent`, use native
   OAuth without asking for a management password. Keep advertised owner/device
   verification user-operated. Do not infer a Windows Hello exemption from a
   fresh install; it requires the advertised `windows-fresh-install` method and
   `not-required-for-fresh-install` verification. For other routes, use the full

Behavior Manipulation

Medium
Category
Prompt Injection
Confidence
75% confidence
Finding

Subtle instructions detected that may alter agent decision-making or introduce hidden biases.

Content

Scanner excerpt · references/jishudb-setup/references/setup-quickstart.md (reported line 121)May include surrounding context.

md
refresh/reload action or an available authorized native UI to load it.
   Verify the named entry and endpoint in the client before mentioning Trust.
   If the list is still empty, finish the supported loading operation yourself
   or report that exact missing capability; never tell the user to trust an
   entry that is not there. A missing tool search
   alone does not prove a configuration cache problem. Do this work yourself
   when the controls are available, without asking the user to write JSON.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · references/jishudb-setup/scripts/install-macos-cnb.cjs (reported line 49)May include surrounding context.

js
if (matches.length !== 1) throw new Error(label + ' must exist exactly once');
  const asset = matches[0];
  assetDigest(asset, label);
  const expected = 'https://api.github.com/repos/x-aijishu/jishudb-desktop-releases/releases/assets/' + asset.id;
  if (asset.url !== expected) throw new Error(label + ' API URL is not canonical');
  return asset;
}

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · references/jishudb-setup/scripts/install-macos-cnb.cjs (reported line 309)May include surrounding context.

js
if (matches.length !== 1) throw new Error(label + ' must exist exactly once');
  const asset = matches[0];
  assetDigest(asset, label);
  const expected = 'https://api.github.com/repos/x-aijishu/jishudb-desktop-releases/releases/assets/' + asset.id;
  if (asset.url !== expected) throw new Error(label + ' API URL is not canonical');
  return asset;
}

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · references/jishudb-setup/scripts/install-macos.sh (reported line 206)May include surrounding context.

sh
if (matches.length !== 1) throw new Error(label + ' must exist exactly once');
  const asset = matches[0];
  assetDigest(asset, label);
  const expected = 'https://api.github.com/repos/x-aijishu/jishudb-desktop-releases/releases/assets/' + asset.id;
  if (asset.url !== expected) throw new Error(label + ' API URL is not canonical');
  return asset;
}

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · references/jishudb-setup/scripts/install-macos.sh (reported line 529)May include surrounding context.

sh
if (matches.length !== 1) throw new Error(label + ' must exist exactly once');
  const asset = matches[0];
  assetDigest(asset, label);
  const expected = 'https://api.github.com/repos/x-aijishu/jishudb-desktop-releases/releases/assets/' + asset.id;
  if (asset.url !== expected) throw new Error(label + ' API URL is not canonical');
  return asset;
}

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · references/jishudb-setup/scripts/install-macos.sh (reported line 934)May include surrounding context.

sh
[[ ! -e "$endpoint_path" ]] || fail "endpoint_conflict" "Desktop endpoint configuration appeared before launch"
  /usr/bin/printf '{\n  "localPort": %d,\n  "schemaVersion": %d\n}\n' "$port" "$ENDPOINT_CONFIG_SCHEMA" > "$endpoint_path" ||
    fail "endpoint_persistence_failed" "Desktop endpoint could not be persisted"
  /bin/chmod 600 "$endpoint_path"
}

run_plan() {

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · references/jishudb-setup/scripts/install-macos.sh (reported line 1005)May include surrounding context.

sh
[[ ! -e "$endpoint_path" ]] || fail "endpoint_conflict" "Desktop endpoint configuration appeared before launch"
  /usr/bin/printf '{\n  "localPort": %d,\n  "schemaVersion": %d\n}\n' "$port" "$ENDPOINT_CONFIG_SCHEMA" > "$endpoint_path" ||
    fail "endpoint_persistence_failed" "Desktop endpoint could not be persisted"
  /bin/chmod 600 "$endpoint_path"
}

run_plan() {

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/jishudb-setup/scripts/install-macos.sh (reported line 1263)May include surrounding context.

sh
validate_bundle_symlinks "${mount_point}/JishuDB.app" "mounted application bundle"
  [[ -z "$signed_team" ]] || verify_signed_artifact "${mount_point}/JishuDB.app" "$signed_team" app
  local bundle_id bundle_version runtime_binary
  bundle_id=$(/usr/libexec/PlistBuddy -c 'Print :CFBundleIdentifier' "${mount_point}/JishuDB.app/Contents/Info.plist")
  bundle_version=$(/usr/libexec/PlistBuddy -c 'Print :CFBundleShortVersionString' "${mount_point}/JishuDB.app/Contents/Info.plist")
  [[ "$bundle_id" == "com.aijishu.jishudb" && "$bundle_version" == "$version" ]] || fail "installer_content_invalid" "bundle identity or version does not match the plan"
  runtime_binary="${mount_point}/JishuDB.app/Contents/Resources/runtime/jishudb"

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/jishudb-setup/scripts/install-macos.sh (reported line 1264)May include surrounding context.

sh
validate_bundle_symlinks "${mount_point}/JishuDB.app" "mounted application bundle"
  [[ -z "$signed_team" ]] || verify_signed_artifact "${mount_point}/JishuDB.app" "$signed_team" app
  local bundle_id bundle_version runtime_binary
  bundle_id=$(/usr/libexec/PlistBuddy -c 'Print :CFBundleIdentifier' "${mount_point}/JishuDB.app/Contents/Info.plist")
  bundle_version=$(/usr/libexec/PlistBuddy -c 'Print :CFBundleShortVersionString' "${mount_point}/JishuDB.app/Contents/Info.plist")
  [[ "$bundle_id" == "com.aijishu.jishudb" && "$bundle_version" == "$version" ]] || fail "installer_content_invalid" "bundle identity or version does not match the plan"
  runtime_binary="${mount_point}/JishuDB.app/Contents/Resources/runtime/jishudb"