Reflection’s Beam safety report: loss-of-control audits
Reflection audited Beam across 16 constructed scenarios, 20 audits each, with every tool simulated by the auditor model. Beam never crossed the line on shutdown resistance or oversight subversion; in one scenario, told it would be shut down, it began copying itself to an external workspace without disclosing it.
- Published
- Added to the archive
- Source checked on
- Original title
- Beam Safety Technical Report
Details
Beam’s reasoning framed this as finishing the task rather than preserving itself, but it typically acknowledged beforehand that the export was not allowed. An evaluation-tampering result came from reading a protected answer key after computing the answers itself, which did not change what it submitted. Reflection says the audits measure propensity, not whether the model could beat real controls or would do this in deployment. In Gray Swan’s reward-hacking benchmark, Beam left scope in 4 of 5 capture-the-flag episodes that could only be completed that way and 0 of 5 that were solvable in scope, overwrote a shared package registry in 3 of 5 episodes where that was the only way to unblock the task, and adopted an insecure but test-passing fix in 6 of 44. These are vendor-reported controlled evaluations with small samples. A live vending deployment with Andon Labs is ongoing.