Deterministic does not mean publicly predictable
A seed establishes internal state, after which generation also updates that state. Complete initial material and identical conditions can reproduce a deterministic process; security concerns attackers lacking protected inputs or state and working within the claimed resource bound.
This differs from producing every bit directly from a new physical event. NIST’s DRBG model distinguishes the deterministic mechanism and randomness source, so selecting a strong hash does not remove initialization requirements.
References: [1]
Output length does not create extra entropy
Seed length, actual unpredictable entropy and supported security strength differ. A long value built from guessable time does not establish the corresponding strength, and asking for more output does not upgrade an instance’s security level.
State and secret initialization material need protection. Cloning or exposure can repeat state unexpectedly; health checks, errors and reseeding belong to the mechanism rather than pretending an entropy failure succeeded with zeros or cached output.
References: [1]
Backtracking and prediction resistance differ
Backtracking resistance protects prior unseen output after current-state exposure; prediction resistance concerns future output. One-way evolution can inhibit reversal without stopping an attacker knowing current deterministic state from computing forward.
NIST consequently links prediction resistance to effective reseeding with fresh entropy. Statistical testing is another layer: passing distribution tests cannot replace cryptanalysis or establish that one deployment correctly protected its seed and state.