Back to skill

Security audit

hallucination-guard

Security checks for vulnerabilities and agentic risk

Overview

The monitor can stop an active agent after a single missing-file check without checking whether interruption is safe.

Layers 0–1 provide basic verification without the automatic termination behavior. Before enabling Layer 3, replace immediate termination with verification and a safe stopping checkpoint. For cross-model audits, choose a provider approved for the source data and exclude secrets or unnecessary private content. The full scanner report completed successfully, and all packaged files were independently inspected.

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
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • 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 (3)

Cross-model auditing forwards files and outputs to another model without a privacy warning.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The workflow directs another model to inspect files and outputs without checking whether that model's provider is authorized to receive their contents. Financial reports, source code, or other sensitive evidence could cross a confidentiality boundary; the verification purpose is legitimate, but does not establish permission to disclose data.

Content

No source excerpt is available for this finding.

The audit workflow shares claims and artifact evidence with another model without a privacy warning.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The template explicitly selects another model and instructs it to read artifacts, potentially placing sensitive claims and tool-returned file contents into that model's context. It prioritizes cost without requiring provider approval, redaction, or disclosure authorization, creating a conditional data-exposure risk rather than evidence of deliberate exfiltration.

Content

No source excerpt is available for this finding.

The monitor automatically terminates agents without warning that interruption may lose in-progress work or leave operations incomplete.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The response protocol directs the monitor to kill an agent based on heuristic alerts, including a single missing artifact, without checking whether interruption is safe. False positives could terminate legitimate work, lose unsaved state, or leave multi-step operations incomplete; a subsequent audit does not restore interrupted operations.

Content

No source excerpt is available for this finding.