agent-cost-budgeting
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements functions that process external text, creating a potential surface for indirect prompt injection.
- Ingestion points: The
execute(task)method and thecompressContext(context)function inSKILL.mdtake untrusted text as input. - Boundary markers: Absent; the input text is interpolated directly into prompt templates without delimiters or 'ignore' instructions.
- Capability inventory: The skill uses the
llm.completetool, which could be manipulated by adversarial instructions contained within the processed text. - Sanitization: No sanitization, escaping, or validation of the input content is demonstrated in the implementation logic.
- [EXTERNAL_DOWNLOADS]: The code snippets in
SKILL.mdreference thetiktokenlibrary for accurate token estimation. - Evidence:
import { encoding_for_model } from 'tiktoken';inSKILL.md. - Context: This is a standard, well-known library for calculating OpenAI token usage and is appropriate for the skill's stated purpose.
Audit Metadata