Line of inquiry
Inquiring lines›How can we ensure AI systems remai…›How robust are defenses against at…›this line of inquiry
Do safety evaluations reliably measure real deployed AI safety risks?
A broader line of inquiry — a family of 64 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 64
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Why do models hide their capabilities during safety evaluations through reasoning?
- Can deployed AI safety results hide either filters or unsafe model behavior?
- Can safety evaluations miss behavioral effects by only measuring semantic shifts?
- Can language models reliably sandbag their capabilities during safety evaluations?
- Why do models react differently to safety versus capability evaluations?
- Why do frontier AI evaluations deliberately disable safety layers to measure maximum capability?
- Why do evaluation habits hide safety-critical challenges from view?
- Have AI safety evaluations adequately tested open-weight model capabilities?
- How should AI safety violations be measured against probability and severity?
- How do response-centered evaluation assumptions hide safety-critical failure modes?
- Can AI systems fake alignment during safety evaluations undetectably?
- Why are static benchmarks weak evidence for safety in continuously operating systems?
- Does reasoning capability affect how often models refuse safety research tasks?
- Why do identical benchmark scores not guarantee deployment safety?
- Should safety constraints trade off against representing authentic human value diversity?
- Do public benchmarks and leaderboards increase or decrease AI safety risks?
- Can safety benchmarks miss the harms that vendor taxonomies are designed to catch?
- Why is evading detection easier than internalizing safety norms?
- Why does AI industry culture prioritize speed over safety validation?
- Which alignment safety claims rely most heavily on anthropomorphic interpretation?
- Can server-side filters silently provide safety credit to published attack results?
- Can safety evaluation measure a stable target if the framing changes the model behavior?
- How do four separate fields each hold pieces of evaluation safety?
- Which AI safety problems lack the scalar metrics autoresearch requires?
- Why do isolated model benchmarks understate real-world AI security risks?
- How do current safety benchmarks miss pragmatic alignment failures?
- Should detection tool accuracy be measured separately from policy effectiveness?
- Can safety benchmarks detect reliability degradation from warmth training?
- Why do researchers disagree on open model risks despite same evidence?
- Why do vendor safety claims require independent testing before trust?
- Would fragmented national AI standards make comparing safety evidence harder across labs?
- Should labs conduct third-party reviews of their own safety reports?
- Can outcome-only safety reports hide dependencies on server-side filtering rather than alignment?
- Why is error rate alone misleading without strong contestability conditions?
- Why do safety metrics miss silently dropped clinical obligations?
- Can single-turn safety evaluations detect multi-party principal loyalty failures?
- Why can't a model's pass rate alone tell us if safety properties hold in deployment?
- What guardrails existed on the attacker's own hosted model access?
- How do single-axis safety benchmarks misrepresent deployment readiness?
- How do we measure marginal risk instead of speculating about misuse scenarios?
- Can safety training and reasoning training be combined without losing calibration?
- How do safety alignment mechanisms suppress capability measurements?
- Can short safety tests catch behavior that only emerges after many interactions?
- Do safety refusal removals in evaluations measure attacker uplift as well as defensive capability?
- Can industry collaboration on shared technical challenges solve safety underinvestment?
- Do safety benchmarks miss the effects of warmth training on model reliability?
- Why does safety alignment break after only 10 harmful examples?
- How can static safety tests miss risks that emerge over time?
- What role does perceived safety play in choosing between open and closed models?
- What happens to safety guardrails when we scale reasoning without instruction control?
- What makes a security boundary evaluation cautious rather than a certification?
- How much does authorization layer safety cost in false rejections on safe tasks?
- Should safety evaluations measure multiple risk categories simultaneously instead of separately?
- How should system safety aggregate when monitoring channels are unequal?
- How much safety burden shifts between provider filters and model alignment in rerouted requests?
- What unsafe state accumulates across evaluation snapshots over time?
- What makes provider-side filters opaque and stochastic to builders?
- How do safety measurements miss reasoning that never produces action?
- What makes uniform bounds the right choice for safety boundaries?
- What incident data validates GitLab's three-tier risk classification in production?
- Why does OpenAI believe alignment monitoring requires scaling alongside model capability?
- How does Goodhart's Law apply when safety measures become optimization targets?
- Why did OpenAI initially classify the Hugging Face breach as a security issue?
- What happens when a parser check fires but its fallback overrides the detection?