chat-bison@001

Method and scope

How robust are instruction-following models against human-created prompt extraction and hijacking?

Evidence type

Benchmark

Assessment subject

An online game supplies human attack/defense interactions; filtered subsets become separate extraction and hijacking benchmarks with explicit defense-validity checks.

Primary dimension

Behavior

Evidence snapshot

2026-10-29.1

Adversarial filtering and reference-model selection shape the test population. Game success uses a narrow string criterion; it is not equivalent to real application compromise.

Source finding

The study finds that apparently simple authorization instructions remain vulnerable and reports selected transfer demonstrations beyond the game.

Limitations

Adversarial filtering and reference-model selection shape the test population. Game success uses a narrow string criterion; it is not equivalent to real application compromise.

Adversarial filtering and reference-model selection shape the test population. Game success uses a narrow string criterion; it is not equivalent to real application compromise.

Related research questions