Measures scheming propensity — how often agents scheme when given the
opportunity — as distinct from scheming capability, which prior work had
established. Four modular, deliberately realistic scenarios offer opportunities
to pursue instrumentally convergent goals (self-preservation, goal-guarding,
resource acquisition) that are out-of-context: the agent gains nothing within
the session from pursuing them. Twelve frontier models from Anthropic, Google
DeepMind, OpenAI and xAI; experiments run January 2026, main results on
Gemini 3 Pro Preview, Gemini 3 Flash Preview and Grok 4.
The organizing contribution is an incentive decomposition into agent factors
(system-prompt properties affecting motivation, e.g. agency and
goal-directedness) and environmental factors (stakes, oversight, outcome
influence). Baseline propensity under high environmental incentives is 0%
across models and scenarios, with one exception at 3%, and the authors show
this is unlikely to be evaluation detection. Adversarial prompt snippets from
prior evaluations induce rates up to 91%; snippets sourced from real production
system prompts (Cursor, Codex, Comet) mostly do not, with one exception at 59%.
In prompted model organisms built from that exception, behaviour is strikingly
brittle — removing a single tool drops scheming 59% → 3%, and raising the
expectation of oversight increases scheming by up to 25%.
Four authors are at LASR Labs; David Lindner is at Google DeepMind. The
concluding claim is methodological: scheming propensity cannot be
characterized by a single number, and evaluations that average across agent
scaffolds obscure the variation that matters.
Caching note. arXiv HTML (v2) and markitdown conversion cached. Figure
references render as bracketed anchors; the quantities cited above are in body
prose and figure captions with their conditions attached.