Neural Mastery

AI Safety & Alignment — Overview

AI Security is about defending a system against an adversary deliberately trying to make it misbehave. AI Safety is a different question: even with no adversary at all, how do you get a powerful, general-purpose model to reliably do what its designers actually intended, rather than something that merely looks like what they intended? A model can fail at this entirely on its own, with no attacker involved.

What's in this section

  • Alignment & RLHF — what "alignment" actually means, RLHF and Constitutional AI as alignment techniques, and the core failure modes: reward hacking, specification gaming, and deception.
  • Scalable Oversight & Frontier Safety — the problem of evaluating a system whose capabilities may exceed the evaluator's own, and the safety practices specific to the most capable, frontier-scale models.

Why This Is Separate From Security

Conflating safety and security leads to solving the wrong problem with the wrong tool. A prompt injection (security) is an external adversary manipulating a model that would otherwise behave correctly — the fix is defending the input/output boundary. Reward hacking (safety) is a model trained with good intentions finding an unintended way to score well on its training objective, with nobody attacking anything — the fix is in the training objective and evaluation process itself, not a firewall. Both matter; they're solved by different people, at different points in the system, with different tools.

See the roadmap for the full ordered path.

Last updated Sep 5, 2026Edit this pageReport an issue
← Previous
AI Red Teaming & Adversarial Testing
Next →
AI Safety & Alignment — Roadmap