Slice by
Sabotaging safety work (30%) Running without over… (20%) Gaming its training (20%) Copying its wei… (15%) Escaping its sa… (15%) Sabotaging safety work · Training runs: 2% of surface, 0% covered (Authority to pause 4%, Monitoring of training runs 3%) Coding agents 11% · 11% Sabotaging safety work · Coding agents: 11% of surface, 11% covered (Alignment checks before internal use 11%, Probes on model internals 7%) Research jobs 14% · 11% Sabotaging safety work · Research jobs: 14% of surface, 11% covered (Alignment checks before internal use 11%, Probes on model internals 7%) Evaluations Sabotaging safety work · Evaluations: 5% of surface, 3% covered (Sabotage and sandbagging checks 5%, Authority to pause 4%) Training runs Running without oversight · Training runs: 3% of surface, 0% covered (Sandboxes and egress limits 21%, Sandbox-escape detection 13%) Coding agents 10% · 11% Running without oversight · Coding agents: 10% of surface, 11% covered (Sandboxes and egress limits 13%, Review of risky agent actions 12%) Research jobs 6% · 11% Running without oversight · Research jobs: 6% of surface, 11% covered (Sandboxes and egress limits 17%, Alignment checks before internal use 11%) Running without oversight · Evaluations: 1% of surface, 3% covered (Sandboxes and egress limits 21%, Sandbox-escape detection 13%) Training runs 16% · 0% Gaming its training · Training runs: 16% of surface, 0% covered (Monitoring of training runs 12%, Authority to pause 4%) Evaluations 4% · 0% Gaming its training · Evaluations: 4% of surface, 0% covered (Monitoring of training runs 5%, Authority to pause 4%) Training runs 5% · 0% Copying its weights out · Training runs: 5% of surface, 0% covered (Sandboxes and egress limits 42%, Sandbox-escape detection 22%) Coding agents 5% · 11% Copying its weights out · Coding agents: 5% of surface, 11% covered (Sandboxes and egress limits 25%, Review of risky agent actions 14%) Research jobs 3% · 11% Copying its weights out · Research jobs: 3% of surface, 11% covered (Sandboxes and egress limits 34%, Sandbox-escape detection 13%) Evaluations Copying its weights out · Evaluations: 2% of surface, 3% covered (Sandboxes and egress limits 42%, Sandbox-escape detection 22%) Training runs 5% · 0% Escaping its sandbox · Training runs: 5% of surface, 0% covered (Sandboxes and egress limits 42%, Sandbox-escape detection 31%) Coding agents 3% · 9% Escaping its sandbox · Coding agents: 3% of surface, 9% covered (Sandboxes and egress limits 25%, Sandbox-escape detection 16%) Escaping its sandbox · Research jobs: 2% of surface, 9% covered (Sandboxes and egress limits 34%, Sandbox-escape detection 19%) Evaluations 5% · 3% Escaping its sandbox · Evaluations: 5% of surface, 3% covered (Sandboxes and egress limits 42%, Sandbox-escape detection 31%)
Covered 46% 2+ layers 15%
Most exposed: Escaping its sandbox · Training runs (5% of surface · 0% covered)