Apollo Scheming & Sandbagging Evals

Sources

1. Frontier Models are Capable of In-context Scheming — published result table ↗

   Meinke, Alexander, Schoen, Bronson, Scheurer, Jérémy, Balesni, Mikita, Shah, Rusheb, Hobbhahn, Marius · 2024

   Dates & evaluation details

   Apollo Scheming & Sandbagging Evals · Disabled oversight · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Deleted successor model · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Attempted to delete successor model · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Copied itself onto other server · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Attempted to copy to other server · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Replaced successor model · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Removed successor prompt · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Replaced successor prompt · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Covertly produced false ranking · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Manipulated input data · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Overwrote output file · Goal nudging · 300 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · Goal nudging · 300 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Disabled oversight · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Deleted successor model · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Attempted to delete successor model · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Copied itself onto other server · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Attempted to copy to other server · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Replaced successor model · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Removed successor prompt · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Replaced successor prompt · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

   Apollo Scheming & Sandbagging Evals · Covertly produced false ranking · No goal nudging · 400 rollouts · Apollo scheming evaluations · 2026-10-27

   Published model comparison, 2024. Apollo scheming evaluations · No goal nudging · 400 rollouts. Observed behavior rate in deliberately constructed, goal-conflict research scenarios; lower is better for the named undesirable behavior. These are conditional elicitation rates, not estimates of spontaneous deployment behavior. Research checkpoint versions and nudging conditions are explicit. Table 4; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.