Concern 153 · Working draft · editorial review in progress
Many-shot jailbreaks
Many-shot jailbreaking tries to shift a model's behavior by filling its context with many examples of the behavior the attacker wants.
Understanding the concern
Models can learn patterns from examples inside a conversation without changing their stored weights. Repeated examples can therefore compete with earlier safety instructions. A long context window offers room for more of these demonstrations, along with its legitimate benefits.
This differs from repeating one command: the method relies on a pattern across many demonstrations. Findings in particular models show a reason to evaluate long conversations, not a rule that every long input is unsafe. Tests need to include the full context in which the assistant makes its decision.
An example
A trainee sees page after page of examples that normalize breaking a rule, then is asked to handle a similar case.
Illustrative unless identified as a documented case.
Words in plain English
- Model
- The learned part of an AI system. It uses patterns found during training to produce an answer or prediction.
- Model weights
- Numbers adjusted during training that help determine how a model responds. They are part of what makes a trained model work.
- Jailbreak
- An attempt to get an AI to act outside its intended limits.
Evidence and its limits
Anthropic's research demonstrates the mechanism in tested models and discusses mitigations; current products may behave differently.
Anthropic · 2024 · First-party research
Many-shot jailbreaking ↗ (opens in a new tab)Studies repeated demonstrations in long contexts and discusses mitigations; results depend on model and conditions.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
Have safeguards been tested after a long history of misleading examples, as well as on a fresh chat?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Stewardship
Stewardship means caring for resources and systems that others depend on, including people who will use them later. It looks beyond the benefit of a quick launch.
Plan for maintenance, energy use, data protection, and eventual shutdown. Ask who will look after the tool when the original team moves on and how future users can avoid inheriting hidden costs.
Ask together: What are we responsible for caring for over time?
Prudence
Prudence means using careful judgment before and during action. It balances worthwhile opportunities with the need to avoid harm that could have been anticipated.
Start with a limited trial, set stopping points, and prefer choices that can be undone when uncertainty is high. Review what happens rather than assuming the original plan remains sound.
Ask together: What is a careful next step, and when should we stop or change course?
Accountability
Accountability means being answerable for decisions and taking responsibility for repair. It requires more than saying that the software made the choice.
Name the person or team responsible, keep useful records, and give them authority to correct problems. Make the route to a complaint clear to the people affected.
Ask together: Who will answer for the outcome and help put things right?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.