Back to skill

Security audit

Phantra AI: Ad Variations

Security checks for vulnerabilities and agentic risk

Overview

The skill creates ad variants with approval before spending, while also exposing changes to future rendering defaults.

Approve each variant's quoted price before rendering. Ask the agent to use explicit model and frame choices for one-off comparisons and change saved defaults only when requested. Model comparisons reuse the source scene rather than creating separate scene IDs. This review covers the packaged instructions, not the connected service's implementation.

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
  • 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)

The model-variation workflow introduces a persistent default-setting capability beyond the requested A/B test.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The model-comparison workflow introduces persistent preference changes without requiring separate consent for their broader scope. If the agent invokes set_preferences during a one-off A/B test, later calls that omit model or frame choices can inherit unintended settings; the quote-and-approval requirements limit immediate financial harm but do not authorize changing defaults.

Content

No source excerpt is available for this finding.

Model variants reuse the existing scene instead of creating the new scene promised by the manifest.

Low
Category
Excessive Agency
Confidence
98% confidence
Finding

The manifest explicitly includes video-model changes among variants created with duplicate_scene, each represented by a new scene. The model workflow instead says no copy is needed and instructs quoting and rendering the same scene with another model, so these variants lack the promised separate scene identity.

Content

No source excerpt is available for this finding.