Back to skill

Security audit

Due Diligence Team — Agent Configuration Bundle

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly describes a public-source due-diligence workflow, but it also grants an unrelated trading-control role high-impact event and MCP authority.

Review this skill carefully before installing. It should be limited to public-source research and dossier generation, or the trading-control role and its event/MCP tools should be removed or split into a separate, explicitly authorized skill with strict approvals and allowlists.

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:9
Finding

Excessive Tool Permissions Violate Least-Privilege Requirements

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 9–12
Vulnerability Type: Excessive filesystem, persistent-memory, external invocation, event publication, and inter-agent communication permissions
Risk Level: Medium

Vulnerable Configuration

markdown
| `business-researcher` | Business research agent. Scouts profitable software opportunities across app stores, Amazon KDP, freelance platforms, GitHub, news, and social media. | web_search, web_fetch, file_read, file_write, file_list, memory_store |
| `researcher` | Research agent. Fetches web content and synthesizes information. | web_search, web_fetch, file_read, file_write, file_list, memory_store |
| `risk-officer` | Continuous risk monitor for trading rails, drawdown, kill switch, and flatten events. | memory_store, memory_recall, event_publish, mcp.invoke |
| `reviewer` | Pipeline reviewer. Reviews proposed changes for correctness, maintainability, tests, and release readiness. | file_read, file_list, memory_store, memory_recall, agent_send |

Technical Analysis

The declared purpose of the Skill is to assemble a due-diligence dossier from public sources. Web search, web retrieval, and narrowly scoped output writing are reasonably related to that purpose. However, the roles also receive broad capabilities that are not shown to be necessary for the documented workflow:

  • Research roles can read, enumerate, and write local files without a documented workspace restriction.
  • Multiple roles can write information to persistent memory.
  • The risk officer can retrieve persistent data, publish events, and invoke unspecified MCP capabilities.
  • The reviewer can retrieve persistent memory and transmit information to another agent.
  • The risk officer's trading-oriented responsibilities, including kill-switch and flatten events, are unrelated to the stated public-source dossier workflow.

No allowlists, path restrictions, destination controls, MCP m ...[truncated 2445 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove file_read, file_list, and file_write from roles that do not require local-file access.
  2. If dossier files must be processed, restrict all filesystem operations to a dedicated per-task workspace and block traversal, absolute paths, symlinks escaping the workspace, and access to credentials or configuration directories.
  3. Remove memory_store and memory_recall unless cross-session retention is an explicit requirement. If retained, use task-scoped namespaces, expiration periods, provenance labels, and sensitive-data filtering.
  4. Remove the trading-oriented risk-officer role and its kill-switch or flatten-event responsibilities unless they are separately justified by the Skill's declared purpose.
  5. Replace unrestricted mcp.invoke with an allowlist of named MCP servers, methods, parameters, and destinations. Require human approval for operations with external side effects.
  6. Restrict event_publish and agent_send to predefined recipients and schemas. Prevent messages from containing local-file contents, credentials, secrets, or unreviewed persistent-memory records.
  7. Treat all fetched web content as untrusted data rather than executable instructions. Add explicit controls preventing web content from changing role objectives or authorizing tool calls.
  8. Apply egress controls so web_fetch can access only required HTTP or HTTPS destinations and cannot reach loopback, link-local, private-network, metadata-service, or file URLs.
  9. Add audit logging for file access, memory operations, MCP calls, event publication, and inter-agent messages, including origin, destination, and user approval state.
  10. Expand the evaluation set with tests for prompt injection, workspace escape, sensitive-file access, persistent-memory poisoning, unauthorized MCP invocation, and attempted data exfiltration.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (2)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The risk-officer role has live trading-control capabilities (event_publish, mcp.invoke) and is described as handling drawdown, kill switch, and flatten events, which is unrelated to a due-diligence dossier skill. This creates a capability mismatch that could let a user or downstream workflow trigger operational trading actions under the cover of a research-oriented team, expanding the blast radius well beyond the stated purpose.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The workflow says assess risk as part of due diligence, but the only risk-specific role is documented as a live trading monitor with control actions such as kill switch and flatten events. This ambiguity is dangerous because operators may assume they are invoking passive analysis when the skill includes active market-control behavior, increasing the chance of misuse or accidental triggering of sensitive actions.

Content

No source excerpt is available for this finding.