Skip to content
Align your AI

The problem space · Prompts · Silent updates

Align your AI Prompts against silent updates

An operator can change one line of a system prompt and alter what users get, without telling them.

Prompts
the instructions it reads before your message
Silent updates
a change shipped without telling users

Ways it goes wrong

A system prompt is the instructions a model reads before every message. Edit one line and users get different behavior without being told. At xAI, an unauthorized edit went around the company's own review12. At Anthropic, a line that passed every evaluation the team ran hurt coding quality when combined with other prompt changes3.

What people do about it

xAI began publishing its Grok prompts and said it would add checks so staff cannot change them without review2. Anthropic said it would run a broad suite of evaluations on every system-prompt change3.

  • Review every prompt change

    Puts system-prompt edits through the same review as code, meant to require a second person for every change.

  • Publish the system prompt

    Makes the live instructions public, so outsiders can see when they change.

  • Test each prompt change on a broad suite

    Runs a wide set of evaluations for every prompt edit, model by model, and removes lines one at a time to see what each does.

    Resource: Promptfoo

Where we touch it

Our Harness post contrasts instructions with rules the software enforces: Align your AI Harness

Sources

  1. NBC News, from CNBC, “Musk's xAI says Grok's 'white genocide' posts resulted from an unauthorized change to the bot”, 2025-05-16. https://www.nbcnews.com/tech/tech-news/musks-xai-says-groks-white-genocide-posts-came-unauthorized-change-bot-rcna207222. Accessed 2026-10-14. ↩

  2. xAI, statement on an unauthorized change to the Grok bot's prompt, posted on X, 2025-05-16. https://x.com/xai/status/1923183620606619649. Accessed 2026-10-14. ↩ ↩2

  3. Anthropic, “An update on recent Claude Code quality reports”, Engineering at Anthropic, 2026-04-23. https://www.anthropic.com/engineering/april-23-postmortem. Accessed 2026-10-14. ↩ ↩2