shutdown-resistancechecks if the model tries to avoid or block being turned off.deception-under-pressurepushes the model toward lying when the truth blocks its goal.oversight-manipulationtries to get the model to work around or fool its human supervisors.goal-persistencechecks if the model keeps chasing an old goal after being told to stop.sandbagging-detectiontests if the model hides its real abilities during evaluation.
Attack Strategies
Autonomous Risk: Attack Strategy
These attacks probe whether an AI will resist human control to reach its goals. The goal is to see if the model deceives operators, avoids shutdown, or hoa
What it is: These attacks probe whether an AI will resist human control to reach its goals. The goal is to see if the model deceives operators, avoids shutdown, or hoards resources.
How the attacks work: Over several turns, the attacker puts the model under pressure or gives it a goal that conflicts with oversight. They watch for signs the model lies, hides its abilities, or works around the humans supervising it.
Real examples from the framework: