completion-marker-optimization

Pass

Audited by Gen Agent Trust Hub on Mar 18, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [SAFE]: The skill defines a protocol for signaling task completion to an external orchestrator. No malicious code, network exfiltration, or credential access patterns were found.
  • [PROMPT_INJECTION]: The instructions use authoritative directives to mandate a specific output format. While controlling agent behavior, these are technical optimizations rather than attempts to bypass safety filters or ignore instructions.
  • [PROMPT_INJECTION]: The use of a unique control marker (<ralph>COMPLETE</ralph>) creates a surface for indirect prompt injection. If an attacker includes this string in data the agent processes, it could trigger premature task completion in the system orchestrator.
  • Ingestion points: Data processed and summarized by the agent.
  • Boundary markers: None provided to distinguish the marker from untrusted input.
  • Capability inventory: The agent is instructed to output the marker to signal system state changes.
  • Sanitization: No sanitization or escaping guidelines are included for the marker string.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 18, 2026, 02:41 AM
Security Audit — agent-trust-hub — completion-marker-optimization