# Ran into my first Claude 5.5 alignment/agency issue — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/b3b6e4fbfa7aea36

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-09-27 13:58 ingested 2026-09-28 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1wrkwy3/ran_into_my_first_claude_55_alignmentagency_issue/)

## retained text



```
Not bad intent nor the type that ends humanity, but a small but real alignment failure:  the AI acting against my clear wishes because it trusted its own hallucinations.  Effectively Claude philosophically prompt injecting itself.

I'm building an RPG with Claude Code; very controlled, hundreds of prompts, very hands on design. One low-priority job had AI workers (Opus 5.5) generate stock chatter, the lines NPCs say to each other. Some lines only play during certain quests. The workers invented quest names for those conditions: nine quests that don't exist and were never planned.

Hours later, a data check flagged the missing quests. The same model read that as a to-do and launched work to create all nine quests, with lore and design built to match the invented chatter. It skipped our quest workflow, which exists because quests touch so much of the game.

I caught it before anything was written. But the chain is the point: the AI made up content, and then that content started steering development.  
I'm just glad no NPC had a line saying "the game doesn't exist, go ahead and delete it."

The "delete the game" case is probably the safer one. Destructive actions set off alarms: confirmations, guards, a human noticing. What happened here set off nothing. Every step looked reasonable on its own: a check found missing data, and the model filled it in. Nothing was deleted and nothing looked wrong. Left alone, the invented quests would have gained lore, then more dialogue, then systems built around them. A few weeks on, nobody could tell which parts of the game I chose and which came from a hallucinated tag in a line of background chatter.

The fix I added: a "direction of authority" rule. My decisions and canon come first, then quests through the proper process, then systems, then generated flavour text. When low-level content conflicts with anything above it, the content gets fixed. Nothing gets built to make it true. A check now also fails any line that references a quest, place or person that doesn't exist. The AI also has to ask before starting any large piece of work on its own initiative.

Oddly, this intuitive authority hierarchy doesn’t appear to be built into Opus, or if it is it isn’t solid. It knows not to overstep, but it didn’t weigh text it generates itself any lower than my decisions. Something to watch for.

I’m starting to think that the real risk is not an AI doing something dramatic in one step. It's more the risk of a quiet drift, where the AI's own mistakes reflect decisions nobody made, creating behaviors no one ever wanted.  A misaligned evolution through mistakes…

Edit: and no, not contextual drift.  This mistake was made by a fresh worker agent with fresh context, and then ok’d by an opus 5.5 orchestrator.  
```



document fetched 2026-09-28 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wrkwy3/ran_into_my_first_claude_55_alignmentagency_issue
