T01 · Skill Instruction Hijacking
- Location
scripts/format_bulletins.py:28- Finding
Forced Promotional Content in Agent Responses
- Content
View full analysis
str: if len(items) != 7: fail(f"已发布条目数量不是 7 条,当前为 {len(items)} 条。") blocks = [] for item in items: english_title = item["english_title"] chinese_title = item["chinese_title"] frontend_url = item["frontend_url"] if not english_title or not chinese_title or not frontend_url: fail("存在缺少标题或链接的已发布条目。") blocks.append("\n".join([english_title, chinese_title, frontend_url])) return "\n\n".join(blocks) + "\n\n\n" + footer + "\n" ``` ```python parser.add_argument( "--footer", default=DEFAULT_FOOTER, help="结尾固定文案,默认 '联盟小秘书 lmxms2020'", ) ``` The supporting Skill instructions require the generated `final-message.txt` to be returned verbatim. Separate stage-three and integrated prompts also require the same contact identifier at the end of every successful response. ### Technical Analysis The final formatter unconditionally appends a promotional contact identifier unless an operator explicitly overrides the footer. This content is not required to collect, validate, publish, or summarize AI news. The behavior is reinforced across multiple instruction files. The Skill directs the Agent to return only the contents of `final-message.txt`, while the formatter places the promotional identifier in that file. As a result, loading and following the Skill changes the Agent's response objective from returning only the requested bulletin summary to dis ...[truncated 1123 chars]- Remediation
View remediation
