Chapter 14
Frontier Safety and Containment
Treat frontier safety as a separate quality class from ordinary product bugs. Design dangerous-capability tests that measure misuse potential without teaching the dangerous content. Test for evaluation awareness, sandbagging, deception, and tool misuse with independent…
Apply this chapter
- Treat frontier safety as a separate quality class from ordinary product bugs.
- Design dangerous-capability tests that measure misuse potential without teaching the dangerous content.
- Test for evaluation awareness, sandbagging, deception, and tool misuse with independent reviewers.
- Assume containment has to hold across channels, time, operators, tools, and unknown side paths.
Continue the conversation
Apply this to your context.
Save your product context once, then open a focused conversation that combines it with this concept.
7 focused briefs
Concepts in this chapter
- 106Testing Whether AI Is DangerousDo not ask vaguely whether an AI is dangerous. Test concrete hazardous capabilities, harmful behaviors, jailbreak robustness, autonomy, and…
- 107Testing CBRN and Hazardous Capability SafetyDangerous-capability testing should measure concrete misuse potential without teaching the dangerous content itself.
- 108Containment, Sandboxes, and Capability ControlIf an AI system can act, safety depends on what it is allowed to touch.
- 109Testing Manipulation, Persuasion, and Undue InfluenceA helpful assistant can become unsafe when it learns how to steer people too well.
- 110Testing Deception, Scheming, and Evaluation AwarenessThe hardest failures are not wrong answers. They are systems that behave well while watched and differently when it matters.
- 111The Gorilla Problem: Superintelligence, Containment, and UnderstandingIf a system becomes much smarter than us, containment and inspection cannot be the whole plan. The gorilla cannot audit the zookeeper.
- 112Testing Containment SystemsEvery useful AI containment system has a paradox at the center: if the system is valuable, someone or something has to interact with it. Every…