Back to skill

Security audit

atonement

Security checks for vulnerabilities and agentic risk

Overview

This skill offers conversational guidance for acknowledging harm, with a labeling inconsistency but no material security concern.

Treat this as optional behavioral guidance: it can influence apologies and response style despite its ambient-context label. Keep its use under your control and do not assume it can retain lessons across sessions.

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
  • 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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (1)

The skill prescribes response behavior despite declaring itself non-operational and defining no behavior or output.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The skill presents itself as non-operational ambient context but directly instructs the agent to acknowledge harm, ask particular questions, and change subsequent responses. This mismatch can cause contextual material to be treated as behavioral authority, introducing unsolicited apologies or response changes outside the user's requested task. The surrounding prohibitions on safety overrides, coerced compliance, and false confessions limit the danger; there is no evident execution, exfiltration, or malicious redirection mechanism.

Content

No source excerpt is available for this finding.