Back to skill

Security audit

evolver-bridge

Security checks for vulnerabilities and agentic risk

Overview

The verdict is benign because the reported findings concern gaps between the skill’s advertised automation and its actual behavior, rather than evidence of intentional harm; confidence is medium because scanning was partial.

The benign verdict does not mean risk-free, and analysis covered only one of three components fully. Before installing, consider that the advertised cycle may require manual steps, the scanner flagged a `git reset --hard` warning, and the skill’s runtime permissions are unknown.

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
Vulnerability Patterns
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (6)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The implementation provides logging and prompt capture for a single legacy run, not the declared end-to-end automation. Its final instruction explicitly leaves solidification as a manual step, directly contradicting the complete-cycle claim. The filesystem writes and subprocess execution support the stated workflow and are not independently suspicious; the mismatch is the materially narrower implemented behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The code serves the stated evolution-orchestration purpose, but materially falls short of the advertised complete end-to-end automation. It explicitly delegates V1 execution and solidification to the user and requires transcript conversion to have happened separately. V2 ingest/distill, retries, logging, and status checks match the description. No unrelated capability is evident in the supplied chunk.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
65% 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 · SKILL.md (reported line 34)May include surrounding context.

md
## Important

- V1 (`evolver run` → `evolver solidify`) is **deprecated** in 2.x — only use when explicitly needed
- `evolver solidify` does `git reset --hard` on failure — commit your work first!
- Logs are appended to `memory/evolution/logs/full-cycle.log`
- All commands retry 3× with backoff; check the log on failure
- **Cross-platform**: scripts use plain `execSync` (POSIX sh) — no PowerShell dependency. They run on WSL/Linux and macOS directly. Do not wrap them in `powershell -Command`; that was the old V1 bug (`powershell: not found` on WSL). `autoexec-manage.js` in the `evolver` skill is the only Windows-only helper (it manages the Windows autoexec daemon via CIM).

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The advertised end-to-end evolution cycle stops after `evolver run` and requires manual solidification.

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

The manifest promises a complete automated V2 ingest/distill or legacy V1 run/solidify cycle with retry. The implementation only invokes evolver run, checks and optionally saves its prompt, and then tells the user to run evolver solidify manually; it implements neither the remaining cycle stages nor retries.

Content

No source excerpt is available for this finding.

The advertised end-to-end automation stops before required steps and does not perform transcript conversion.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The manifest promises a complete automated cycle, including legacy V1 run/solidify and transcript conversion. However, V1 ends by instructing the user to execute the plan and run evolver solidify manually (L129), while V2 exits when no converted transcript exists and tells the user to run session-to-material first (L137–L142). These are explicit manual prerequisites or follow-up steps, contrary to the claimed end-to-end automation.

Content

No source excerpt is available for this finding.