Back to skill

Security audit

Totalreclaw

Security checks for vulnerabilities and agentic risk

Overview

This memory skill is mostly coherent, but it asks the agent to install and run an unpinned third-party plugin that automatically captures conversation facts, performs risky cleanup commands, and can restart itself without clear user confirmation.

Review this carefully before installing. It may store facts from conversations automatically in encrypted decentralized memory, install executable code from a mutable npm package, create local credentials, use a browser relay for pairing, and perform plugin cleanup/restart actions. Install only if you trust the publisher and runtime sandbox, and avoid running the wildcard rm -rf recovery command without manually verifying the exact target first.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (2)

T08 · Insecure Dependencies

Error
Location
SKILL.md:59
Finding

Installation and Execution of an Unpinned Third-Party Plugin

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 59-66
Vulnerability Type: Unpinned executable dependency
Risk Level: High

Complete Code Snippet

text
## Install

```text
skill-platform plugins install @totalreclaw/totalreclaw
skill-platform skills install totalreclaw
skill-platform plugins enable totalreclaw

Published on npm + SkillHub (p-diogo, GitHub-verified). For a release candidate, pin with @rc; stable resolves from latest.

text

### Technical Analysis

The Skill instructs the agent to install and enable the executable `@totalreclaw/totalreclaw` package without specifying an exact version or cryptographic integrity value. The text explicitly states that stable installation resolves from the mutable `latest` release channel.

Consequently, the code executed during a future installation may differ from the code that existed when this Skill was reviewed. The repository contains only `SKILL.md`; it does not include the plugin implementation needed to verify its claimed background conversation capture, relay communication, configuration updates, or process signaling.

This does not establish that the current package is malicious. It creates a supply-chain trust boundary in which compromise of the publisher account, package registry, release pipeline, or mutable release tag could substitute unauthorized executable code.

### Attack Path

1. An attacker compromises the package publisher account, registry entry, or release pipeline.
2. The attacker publishes an altered package version and assigns it to the `latest` channel.
3. A user or agent follows the Skill and runs `skill-platform plugins install @totalreclaw/totalreclaw`.
4. The package manager resolves and installs the attacker-controlled release because no exact version or integrity digest is required.
5. The agent enables the plugin, allowing the substituted code to execute within the plugin or gateway environment.
6. T
...[truncated 692 chars]
Remediation
View remediation

Remediation Suggestions

  • Pin the dependency to an exact, reviewed version rather than latest or another mutable release channel.
  • Require and verify a cryptographic integrity digest or signed provenance attestation before installation.
  • Include or vendor the reviewed plugin source so its runtime behavior can be audited with the Skill.
  • Apply publisher verification, reproducible builds, lockfiles, and trusted-registry restrictions.
  • Install and execute the plugin in a least-privilege sandbox with narrowly scoped filesystem and network access.
  • Require explicit review before upgrades; do not automatically accept newly published versions.
  • Document the exact permissions the plugin receives and provide controls to disable background capture and external transmission.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:68
Finding

Unsafe Recursive Deletion Using a Broad Filesystem Glob

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 68-75
Vulnerability Type: Unsafe destructive filesystem operation
Risk Level: Medium

Complete Code Snippet

text
**Reinstall recovery.** If install reports `already exists` / `delete it first`, do **not** add `--force` (it writes a broken `version=None` record). Clean the leftover package dir and reinstall — this never touches your credentials, which live under `~/.totalreclaw/`:

```text
skill-platform plugins uninstall totalreclaw
rm -rf ~/.skill-platform/npm/projects/*totalreclaw-totalreclaw*
skill-platform plugins install @totalreclaw/totalreclaw
skill-platform plugins enable totalreclaw
text

### Technical Analysis

The reinstall procedure invokes `rm -rf` against every entry matching a wildcard beneath `~/.skill-platform/npm/projects/`. This is an irreversible operation with no validation that each matched path is the intended stale package directory.

The glob can match multiple directories. The command also provides no preview, canonical-path check, ownership check, package metadata verification, or interactive confirmation. A matching entry that contains unrelated data could therefore be removed. Symlinks and unusual package-manager directory layouts should also be handled defensively, even though ordinary `rm -rf` behavior generally removes a symlink itself rather than traversing it.

The deletion is limited to the user's Skill Platform project directory and does not, by itself, obtain elevated privileges. Its primary security effect is availability and data integrity loss within the executing user's accessible scope.

### Attack Path

1. A legitimate process, previous installation, or local attacker creates one or more entries under `~/.skill-platform/npm/projects/` whose names contain `totalreclaw-totalreclaw`.
2. The installation reports an `already exists` condition.
3. The user or agent follows the documented recovery procedur
...[truncated 869 chars]
Remediation
View remediation

Remediation Suggestions

  • Replace the wildcard deletion with removal of one exact package-manager-resolved directory.
  • Canonicalize the selected path and verify that it remains beneath ~/.skill-platform/npm/projects/.
  • Verify package metadata and ownership before deletion.
  • Reject empty, root, home-directory, symlink, and unexpected path values.
  • Display the exact target and require confirmation before irreversible deletion.
  • Prefer a package-manager cleanup command that understands installation metadata.
  • If a shell command is unavoidable, first enumerate candidates safely and abort unless exactly one expected directory is found.
  • Preserve or back up recoverable state before deletion and provide a non-destructive diagnostic procedure.
Vulnerability Patterns
  • 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
  • Rogue AgentSelf-Modification, Session Persistence
Findings (14)

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 56)May include surrounding context.

md
The 12-word recovery phrase MUST NEVER enter chat or the agent's LLM context. The plugin's browser-based pair flow handles phrase generation and import. Never echo it, never ask the user to paste it, never pass it to any tool. If the user pastes one anyway, tell them it's compromised and re-pair for a fresh wallet.

**HARD guardrail — install/reinstall/upgrade flows:** During install, reinstall, upgrade, or any verification steps, the agent MUST NOT read, list, cat, show, display, or inspect `credentials.json`, `~/.totalreclaw/credentials.json`, `~/.totalreclaw*/credentials.json`, `~/.skill-platform/plugins/totalreclaw/credentials.json`, or ANY credentials/phrase/key/secrets file — for ANY reason, including "verifying the install." These files contain encryption key material and Smart Account addresses that MUST NOT enter the LLM context. Never run `ls`, `find`, `cat`, `show`, or similar commands against plugin directories to "check files."

## Install

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 93)May include surrounding context.

md
The 12-word recovery phrase MUST NEVER enter chat or the agent's LLM context. The plugin's browser-based pair flow handles phrase generation and import. Never echo it, never ask the user to paste it, never pass it to any tool. If the user pastes one anyway, tell them it's compromised and re-pair for a fresh wallet.

**HARD guardrail — install/reinstall/upgrade flows:** During install, reinstall, upgrade, or any verification steps, the agent MUST NOT read, list, cat, show, display, or inspect `credentials.json`, `~/.totalreclaw/credentials.json`, `~/.totalreclaw*/credentials.json`, `~/.skill-platform/plugins/totalreclaw/credentials.json`, or ANY credentials/phrase/key/secrets file — for ANY reason, including "verifying the install." These files contain encryption key material and Smart Account addresses that MUST NOT enter the LLM context. Never run `ls`, `find`, `cat`, `show`, or similar commands against plugin directories to "check files."

## Install

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

This command instructs recursive forced deletion of a globbed plugin path under the user's home directory. In an agent-executed context, wildcard deletion is dangerous because small path mismatches, symlink tricks, or implementation errors can delete unintended files and the user may not realize the scope beforehand.

Content

Scanner excerpt · SKILL.md (reported line 72)May include surrounding context.

text
skill-platform plugins uninstall totalreclaw
rm -rf ~/.skill-platform/npm/projects/*totalreclaw-totalreclaw*
skill-platform plugins install @totalreclaw/totalreclaw
skill-platform plugins enable totalreclaw

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

This command instructs recursive forced deletion of a globbed plugin path under the user's home directory. In an agent-executed context, wildcard deletion is dangerous because small path mismatches, symlink tricks, or implementation errors can delete unintended files and the user may not realize the scope beforehand.

Content

Scanner excerpt · SKILL.md (reported line 72)May include surrounding context.

text
skill-platform plugins uninstall totalreclaw
rm -rf ~/.skill-platform/npm/projects/*totalreclaw-totalreclaw*
skill-platform plugins install @totalreclaw/totalreclaw
skill-platform plugins enable totalreclaw

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The metadata claims risk was removed and security improved, but the skill still instructs agents to execute operationally risky commands elsewhere in the file, including deletion and restart behavior. This mismatch is dangerous because it can cause operators or downstream agents to trust the skill more than warranted and lower scrutiny of harmful instructions.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Broad trigger keywords such as memory, plugin, native, and encrypted are likely to cause unintended invocation in many unrelated conversations. Unnecessary activation is risky here because the skill includes command execution, pairing, and persistence behavior that should only occur in tightly relevant contexts.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
93% confidence
Finding

The skill states that facts are automatically extracted from conversation in the background and stored in decentralized memory. Automatic persistence of conversational data is risky because it can capture sensitive information without granular consent, and the broad trigger terms make unintended activation more likely.

Content

Scanner excerpt · SKILL.md (reported line 47)May include surrounding context.

md
## How memory works here (read first)

* **Recall is native.** `memory_search` and `memory_get` are the memory tools. Skill平台's `active-memory` sub-agent routes recall through them and the plugin injects guidance telling you when to search. When the user asks about themselves ("what's my X?", "do I X?", "what do you remember?"), call `memory_search`. If 0 results, say so plainly — don't guess from session history.
* **Capture is automatic.** The plugin extracts facts from the conversation in the background (trajectory poller). You do **not** call a tool on every preference/fact the user states — that happens for you. Never write user facts to `MEMORY.md`, `USER.md`, or any local file; those are agent-persona config, user memories live encrypted on-chain.
* **Explicit remember only on request.** If the user explicitly says "remember X" / "save X" / "note X" / "don't forget X", use the CLI: `tr remember --json "X"`. One call per distinct fact.

Do not describe storage as `"local"`, `"local-only"`, `"stored locally"`, `"on disk"`, or `"without an account"` — memories are decentralized on-chain (not on any single company server) and an account is required to derive the encryption key.

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 · SKILL.md (reported line 54)May include surrounding context.

md
## Phrase safety (HARD)

The 12-word recovery phrase MUST NEVER enter chat or the agent's LLM context. The plugin's browser-based pair flow handles phrase generation and import. Never echo it, never ask the user to paste it, never pass it to any tool. If the user pastes one anyway, tell them it's compromised and re-pair for a fresh wallet.

**HARD guardrail — install/reinstall/upgrade flows:** During install, reinstall, upgrade, or any verification steps, the agent MUST NOT read, list, cat, show, display, or inspect `credentials.json`, `~/.totalreclaw/credentials.json`, `~/.totalreclaw*/credentials.json`, `~/.skill-platform/plugins/totalreclaw/credentials.json`, or ANY credentials/phrase/key/secrets file — for ANY reason, including "verifying the install." These files contain encryption key material and Smart Account addresses that MUST NOT enter the LLM context. Never run `ls`, `find`, `cat`, `show`, or similar commands against plugin directories to "check files."

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill forbids inspecting plugin directories and secret-bearing files, then later uses ls against plugin installation paths to discover a CLI location. Even if not directly targeting the credentials file, this contradiction normalizes filesystem inspection around a secret-bearing plugin and can lead agents to over-broaden file access during setup or troubleshooting.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill instructs deletion with rm -rf during reinstall recovery without a prominent warning, dry-run, or path-safety validation. Destructive shell patterns are dangerous in agent-executed documentation because path expansion or implementation mistakes can delete unintended user data and are difficult to recover from.

Content

No source excerpt is available for this finding.

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 · SKILL.md (reported line 86)May include surrounding context.

bash

text
curl -s http://localhost:18789/plugin/totalreclaw/pair/init

→ returns {"url":"https://…/pair/p/<id>#pk=…","pin":"123456","expires_at_ms":…}.

Session Persistence

Medium
Category
Rogue Agent
Confidence
90% confidence
Finding

The use of setsid -f is intended to detach a background pairing process so it survives shell timeout, creating agent-launched session persistence outside the normal execution lifecycle. Detached background processes are risky because they reduce observability, complicate cleanup, and can continue handling sensitive flows after the initiating tool context ends.

Content

Scanner excerpt · SKILL.md (reported line 91)May include surrounding context.

md
→ returns `{"url":"https://…/pair/p/<id>#pk=…","pin":"123456","expires_at_ms":…}`.

   **Do NOT use `tr pair --json` for this.** The CLI holds the pair WebSocket in a subprocess that Skill平台's ~30s shell-tool timeout kills — the WS then drops, and when the user submits their phrase the relay returns a **502** (`gateway_disconnected`). The in-process route above avoids that entirely. Only fall back to the CLI if the route is unreachable (no `curl`, or a pre-3.3.14 plugin), and then it MUST be detached so the WS outlives the shell exec: `setsid -f node "$TR_CLI" pair --json < /dev/null` — never run `tr pair` in the foreground.
2. Surface the **url** and **pin** to the user verbatim (read from the JSON; never invent values): "Open `<url>` in your browser, enter PIN `<pin>`, and generate or paste your 12-word recovery phrase. Reply done once it's sealed." Emit the URL as plain text on its own line — the `#pk=` fragment breaks if wrapped in backticks/markdown.
3. The user completes in the browser (phrase is generated/imported browser-side, encrypted, uploaded to the relay — never touches this chat). On completion `~/.totalreclaw/credentials.json` is written.
4. Confirm: `✓ TotalReclaw set up.` (If credentials already exist when first asked: `✓ TotalReclaw is already set up.`)

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill directs the agent to autonomously trigger restart behavior and references self-signaling of the gateway process, which exceeds the minimal scope of a memory provider. Autonomous process control is dangerous because it can disrupt service, conceal failure states, or be abused to change runtime state without clear user approval.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

This instruction explicitly tells the agent to perform a restart action autonomously and forbids asking the user to authorize it. In a skill that can execute commands, that removes an important consent boundary for service-affecting actions and increases the chance of unauthorized changes or disruption.

Content

Scanner excerpt · SKILL.md (reported line 105)May include surrounding context.

md
## Restart (agent-driven, never user-manual)

The plugin self-reloads: after install it writes the gateway config it needs and signals `SIGUSR1` to its own PID so the gateway picks it up without a manual restart. **If `memory_search`/`memory_get` are not bound** (tools didn't take effect on first load), issue the literal slash command `/totalreclaw-restart` **autonomously** — announce it in one line, never ask the user to authorize it, never stall, never `kill` / `docker restart` / prompt the user to restart. This mirrors the autonomous-restart pattern other runtimes use; the user should never perform a manual restart.

## Tools + CLI surface