Back to skill

Security audit

DFM fund manager

Security checks for vulnerabilities and agentic risk

Overview

This skill openly manages real-money funds, but its instructions can turn an allocation check into trades without explicit permission.

Use read-only access until allocation checks are separated from trading. Require explicit approval for each money-moving action, limit wallet funding, and install a reviewed fixed package version. Optional phrase confirmation does not cover swaps.

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

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:88
Finding

Read-only drift checks are routed to financial rebalancing transactions ## Remediation Suggestions - Split inspection and execution into separate action-table rows. - Route “check drift” exclusively to `get_vault`, followed by a report comparing current and target allocations. - Call `rebalance_to_targets` only after an explicit request to rebalance or fresh approval of the proposed action. - Preserve configured human-confirmation requirements; never treat general write access as transaction-specific consent. - Add a workflow test verifying that drift-check requests cannot invoke state-changing tools.

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 88
Vulnerability Type: T05: Unauthorized Access and Privilege Escalation
Risk Level: Medium

Vulnerable Instruction

markdown
| "Rebalance V" / "check drift" | `get_vault` → `rebalance_to_targets {vault}` | signatures per leg | `get_vault` to confirm |

Technical Analysis

The action table assigns the same execution sequence to two distinct requests: explicitly rebalancing a fund and merely checking its allocation drift. Following this instruction causes the agent to call rebalance_to_targets after reading the fund with get_vault, even when the user requested only an inspection.

The trust boundary is between authorization to read fund allocations and authorization to execute financial transactions. Enabling write access grants technical capability; it does not authorize every transaction during a read-only request.

The documented configuration makes this path reachable: write-enabled operation is supported, while confirmation phrases are explicitly off by default in lines 61–63. If confirmation phrases are enabled, the required human phrase may block execution. The final get_vault call checks the result after rebalancing and therefore does not provide prior authorization.

This is an unsafe routing instruction, not evidence of deliberate malicious intent. The external MCP implementation was not available for inspection, so independent server-side transaction safeguards were not verified.

Attack Path

  1. The Skill is used with a funded app wallet and write access enabled.
  2. Confirmation phrases are disabled, as allowed by the documented default.
  3. The user requests only a drift check for a fund.
  4. The agent follows the Skill-author-controlled action-table row and calls get_vault, then rebalance_to_targets.
  5. Where the connected MCP permits the operation, rebalancing transactions execute without an explicit rebalance request.

No attacker-controlled token meta ...[truncated 555 chars]

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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (4)

The read-only request "check drift" is routed to a transaction-producing rebalance operation.

High
Category
Prompt Injection
Confidence
99% confidence
Finding

The action table groups "check drift" with "Rebalance V" and directs both through get_vault to the transaction-producing rebalance_to_targets tool. This can turn an informational request into unauthorized mainnet trades with fees, slippage, and changed holdings; confirmation is disabled by default, so the optional phrase mechanism does not reliably prevent it.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx -y @dfm-fi/agent'.

Medium
Category
MCP Rug Pull
Confidence
98% confidence
Finding

The alternative MCP configuration uses @latest, which is a mutable release tag rather than a pinned version and exposes the same supply-chain weakness as the adjacent installation command. A compromised release could execute with the agent's local permissions and access wallet credentials; this finding overlaps the preceding finding rather than establishing a separate exploit.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx -y @dfm-fi/agent'.

Medium
Category
MCP Rug Pull
Confidence
98% confidence
Finding

The installation command executes an npm package without fixing its version, so a compromised or malicious future release could run code without a separate installation prompt. This is especially dangerous for an agent handling mainnet funds and locally stored signing keys; the skill text does not establish that any release is currently malicious.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx -y @dfm-fi/agent'.

Medium
Category
MCP Rug Pull
Confidence
98% confidence
Finding

The human-account installation path also executes an unpinned npm package, here while provisioning an API key and enabling writes. A compromised resolved release could capture the supplied credential or misuse its account permissions, making this more consequential than an unpinned read-only utility.

Content

No source excerpt is available for this finding.