Back to skill

Security audit

Evomap Evolver

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed self-evolution tool, but it gives network-sourced instructions and automatic update/reporting paths too much influence over an agent that can change code.

Install only in a tightly scoped repository with review mode, Hub reuse disabled or restricted to trusted assets, auto-update disabled, and no broad GitHub token unless you want automatic issue creation. Avoid loop or worker mode until you are comfortable with remote tasks, persistent identity, and git rollback behavior.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
src/gep/prompt.js:12
Finding

Untrusted Hub assets are injected into authoritative executor instructions

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (198)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The code does include behavior consistent with a self-evolution engine: it runs evolve.run(), supports looped evolution, pending solidify state, and a solidify workflow. However, the supplied chunk exposes several substantial capabilities not reflected in the declared description. In addition to evolution, it provides a fetch command that registers with a remote hub and downloads skills over the network, a review command that inspects git diffs and can approve or reject changes with rollback, a distill command that ingests an external response file to produce genes, heartbeat communication to a hub, and asset-log reporting. These are not merely internal implementation details of evolution; they are user-facing operational capabilities with distinct purposes. The declared permissions (network, shell) are compatible with some of this behavior, so permissions are not the issue. The mismatch is that the description materially understates the breadth of the tool’s actual functionality.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The code does not perform self-evolution logic. It reads existing assets and events from storage, filters eligible capsules/genes/events, computes schema/asset IDs for events if needed, formats them as JSON or A2A protocol messages, prints them to stdout, and optionally sends them via a transport when --persist is used. While it touches evolution-related data (e.g., EvolutionEvent), its primary purpose is export/publication, not analyzing runtime history to identify improvements or applying protocol-constrained evolution. Therefore the declared description materially misrepresents the actual behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description suggests a system that reviews an agent’s runtime history and performs constrained self-improvement. The code does none of that. Its primary function is ingesting external A2A assets from stdin/file, validating integrity, downgrading confidence, storing candidates, recording them in a memory graph, and optionally sending decision messages. This is a materially different purpose: external content intake and quarantine workflow rather than agent self-evolution. The declared permissions are not the issue here; the mismatch is in core behavior and purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The code does not analyze agent runtime history or derive improvements. Instead, it parses command-line arguments, looks up an externally sourced candidate by type/id, verifies a --validated flag, performs a narrow safety check for Gene validation commands, marks the asset as promoted, writes it into local stores, and optionally sends accept/reject decision messages. This is materially different from a self-evolution engine. While the optional A2A decision emission loosely relates to protocol handling, the core behavior is a manual promotion workflow for external assets, not runtime-driven agent self-improvement.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The code is a reporting/analysis utility, not an engine that evolves AI agents. It reads evolution_history_full.md, extracts interesting entries, groups them by skill, and writes evolution_detailed_report.md. There is no code that modifies agent behavior, applies improvements, enforces evolution protocols, uses network access, or invokes shell commands. While analyzing history is loosely related to the description, the primary declared purpose—identifying improvements and applying protocol-constrained evolution—is materially overstated relative to the actual behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description says this skill is a self-evolution engine for AI agents that analyzes runtime history and applies constrained evolution. The supplied code does none of that. Its primary purpose is packaging and release preparation: it reads a public manifest, builds a dist/public-style directory, applies file rewrites, rewrites package scripts, derives a suggested SemVer version from git log subjects, and writes build/versioning notes to local files. While shell access is declared and used for git log commands, the behavior is materially different from the stated purpose. There is no runtime agent-history analysis, no improvement synthesis, and no protocol-constrained self-modification/evolution logic.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents this skill as an active self-evolution engine for AI agents. However, the supplied code only performs offline file processing: it reads a specific local log, extracts embedded Feishu send-command data, converts timestamps, deduplicates entries, and writes a Markdown history report. It does not analyze runtime behavior for improvements, modify agent protocols, apply any evolution, or use shell/network capabilities. This is a materially different primary purpose, so the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description describes an active self-evolution system for AI agents with analysis and change-application behavior. The supplied code only reads git commit history from the local repository, filters commits by the term "Evolution," reformats the results, and saves a markdown report. It does not inspect runtime history, evaluate improvements, modify agent behavior, or enforce any protocol-constrained evolution. While shell access is used for git log, that is consistent with implementation details; the mismatch is in the core purpose and actual functionality.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The code is a standalone report generator. It reads JSON/JSONL files from repository memory/assets directories, computes aggregate statistics for EvolutionEvent entries keyed by personality state, ranks them, and prints a summary. It does analyze historical runtime/evolution data, which partially overlaps with the description, but it does not perform the core claimed behavior of a 'self-evolution engine': there is no mutation selection, protocol-constrained evolution logic, state modification, or application of improvements. The declared permissions are over-broad rather than dangerous here, since the script only uses local filesystem access. Because the actual primary purpose is reporting on past evolution/personality outcomes rather than evolving the agent, this is a material description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description suggests an active self-evolution system for AI agents: analyzing runtime history, determining improvements, and applying constrained evolution actions. The actual code does none of that. It simply reads evolution_history_full.md, parses entries, heuristically categorizes them, extracts summary lines, and writes a markdown report to evolution_human_summary.md. There is no network or shell usage, no agent modification, no decision-making about improvements, and no mechanism that applies any evolution or protocol constraints. This is a materially different primary purpose: human-facing summarization/reporting rather than self-evolution.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The code does not analyze runtime history, perform agent self-improvement, or apply protocol-constrained evolution. Its primary function is CI/release engineering: checking git cleanliness and branch, cloning a public repo, copying build artifacts, committing/pushing changes, tagging releases, creating GitHub releases, and publishing to ClawHub. While the declared permissions of network and shell are consistent with the implementation, the declared purpose is materially inaccurate because the actual behavior is a software publishing pipeline rather than an AI self-evolution engine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The supplied code chunk does not implement runtime-history analysis or protocol-constrained self-evolution. Instead, it serves as an operational recovery utility: it delays, searches for an evolver entrypoint in the workspace, and restarts that program. While this may support an evolver system, the primary behavior of this chunk is restart orchestration rather than self-evolution logic. The shell and filesystem usage are consistent with that restart role, but that role is materially different from the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The code does not implement any self-evolution engine behavior. It neither analyzes AI agent runtime history nor identifies behavioral improvements nor applies any evolution or protocol-constrained changes. Instead, it is a release automation utility: it inspects git commit messages, determines whether to bump major/minor/patch semver, and saves the suggestion. While shell access is used via git commands, that supports repository version analysis, not the declared AI-agent evolution purpose. This is a material purpose mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description says this skill is a self-evolution engine for AI agents, implying analysis of execution history and application of constrained improvements. The supplied code does none of that. It is a simple JavaScript utility for validating module exports by requiring modules passed on the command line and checking that exports are present and non-empty. Its primary purpose is materially different from the declared purpose. The trigger profile also differs: the description lists no triggers, while the code is explicitly invoked as a CLI script with module paths as inputs. Although the declared permissions include network and shell, the code itself does not use network access and only behaves as a local validation script. This is a clear description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description says this skill is a self-evolution engine for AI agents, implying analysis of runtime behavior and modification/evolution logic. The supplied code instead only runs node --test against matching test files, parses the output for pass/fail counts, and reports success or failure. It does not inspect runtime history, perform optimization, modify agents, or enforce evolution protocols. The code's actual purpose is validation/testing, which is materially different from the declared purpose. The declared permissions include network and shell; shell use is consistent with running tests, and over-declared permissions alone are not the reason for the mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The declared description emphasizes a self-evolution engine that analyzes runtime history and applies protocol-constrained improvements. While there is a small overlap in reading evolution events and computing a capsule success streak, the primary behavior of this file is not performing self-evolution or applying improvements. Instead, it is an A2A exchange/helper module for importing, validating, annotating, and exporting specific asset types (Gene, Capsule, EvolutionEvent). It parses external JSON inputs, unwraps protocol messages, determines whether assets are eligible for broadcast, and reads local files. The filesystem access is also not suggested by the description, and the agent-to-agent messaging/export behavior is a materially different purpose from runtime-history-driven self-evolution.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The code chunk does not analyze runtime history or perform self-evolution logic. Its primary purpose is infrastructure for communication: defining A2A message types, building publish/fetch/report/decision/revoke messages, managing node IDs and shared secrets, persisting them on disk, using file and HTTP transports, registering with a hub, sending periodic heartbeats, and polling for events/work. While this may support a larger evolution system, this specific code is materially different from the declared purpose. The declared permissions of network and shell are not the key issue; rather, the behavior is primarily protocol/transport and telemetry, not self-improvement analysis or evolution application.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied code does not implement a self-evolution engine or runtime-history improvement analysis. It only provides helper functions to append asset call records to a local JSONL log, read them back with filters, and summarize counts for display. While this may be a supporting component within a larger evolution system, this code chunk’s primary behavior is asset-call logging/reporting, which is materially narrower and different from the declared purpose. There is no evidence here of network use, shell execution, agent self-modification, runtime analysis, or protocol-constrained evolution logic.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied code does not implement runtime-history analysis, self-evolution, or protocol-constrained agent improvement. Instead, it is a small asset utility module focused on JSON/string preview formatting and adding schema/version and computed IDs to asset objects. While this could be a supporting utility within a larger system, the code chunk itself has a materially different primary purpose from the declared description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The code chunk does not implement an AI self-evolution engine. It contains helper functions for directory creation, timestamping, text clipping, writing prompt artifacts and metadata files to a local memory directory, and formatting a sessions_spawn(...) invocation string. While declared permissions include network and shell, this snippet uses neither; instead it uses filesystem access. The primary purpose is materially different from the description because there is no analysis of runtime history, no identification of improvements, and no application of protocol-constrained evolution.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied code does not analyze runtime history, propose improvements, or evolve agent behavior. Instead, it provides deterministic object serialization and SHA-256-based content hashing/verification for assets. This is a materially different primary purpose from the declared self-evolution engine. While the declared permissions include network and shell, this code uses neither; over-declared permissions alone are not the issue, but they further indicate the description does not match this module's actual function.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The code chunk does not perform runtime-history analysis, improvement discovery, or any self-evolution behavior. Its primary purpose is device fingerprinting and stable node identity generation/persistence. While some of the accessed resources could be incidental in a larger system, this specific module is materially different from the declared skill description. The declared permissions include network and shell; the code mostly uses filesystem and OS metadata access, with a limited shell-like subprocess call on macOS (ioreg), but the main issue is purpose mismatch rather than over-declared permissions.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The declared purpose describes an AI self-evolution engine focused on analyzing runtime history and applying constrained improvements. The supplied code does not implement history analysis, agent improvement logic, or protocol-constrained evolution. Instead, it gathers environment metadata (device ID, Node version, platform, architecture, OS release, hashed hostname, hashed cwd, package name/version, region, container status, timestamp) and computes comparison keys for grouping environments. While such telemetry could theoretically support a larger evolution system, this chunk’s primary behavior is environment fingerprinting and host classification, which is materially different from the declared purpose and includes system identification capabilities not reflected in the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The code chunk does not analyze runtime history or perform agent self-evolution logic. Its primary purpose is repository state management and rollback safety around git-controlled changes. While such utilities could support a larger evolution system, this specific code is materially different from the declared description and exposes concrete file/git manipulation behavior not reflected in the stated purpose. The declared permissions are not a problem by themselves, but the actual implemented behavior here is operational git/file rollback tooling rather than an evolution engine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The description says this skill is a self-evolution engine focused on analyzing runtime history and applying constrained evolution. This code does not implement evolution logic, runtime-history analysis, or protocol-constrained self-modification. Instead, its primary function is auxiliary Hub asset review submission: it builds review content, computes ratings from outcome data, sends POST requests to /a2a/assets/:assetId/reviews, persists a local reviewed-asset history file, and logs review-related events. While this may relate to a broader evolution system, this chunk’s actual purpose is materially different from the declared description, so it should be flagged as a mismatch.

Content

No source excerpt is available for this finding.