The problem of determining what dose levels of an agent are safe, e.g., non-carcinogenic, cannot be resolved unless one first defines some level of permissible risk, no matter how small, rather than insisting on absolute safety. Both because of practical considerations and statistical variation, the determination of low-risk dose levels, for example 1/100 million, cannot be made directly but must be by extrapolation from observed data. A conservative approach for doing so is given. In addition to an arbitrary definition of “virtual safety,” it is necessary to define an arbitrarily high statistical assurance level and a rule for extrapolation by use of an arbitrarily shallow slope. Illustrative data by Bryan and Shimkin (J. Nat. Cancer Inst. 3: 503–531, 1943) on the carcinogenic action of methylcholanthrene yield a “safe,” 1/100 million, dose of 9 × 10−8 mg per mouse when a statistical assurance level of 99 percent and a conservative probit slope of 1 normal deviate per log for extrapolation are used. The principles given are of general applicability in other safety-testing problems, the point of emphasis being that since direct observation cannot be made that the risk at some dose level is clearly low, indirect conservative procedures for the determination of low risk levels must be made. The arbitrary risks and definitions for so doing may change with circumstances. The procedure does not require specification of an experimental protocol; the “safe” dose is determined on the basis of whatever data are available. Minimum protocols may, however, be desirable, since greater amounts of data will ordinarily permit specifying large “safe” levels.