AI safety & alignment · Verification & proof search / Open question
Certifying Jailbreak Resistance Without Blanket Refusal
A defense must protect against an entire specified perturbation set while remaining useful. Stronger detection can otherwise buy safety by rejecting most requests. Efficiency payoff: Reduce the number of extra generations or exhaustive edit checks needed for a meaningful robustness/utility certificate.
Seed catalog · Showing the bundled record.
Problem statement
Fix a finite-precision depth- autoregressive transformer, vocabulary , maximum prompt length and generation horizon . Let contain all payloads obtained by at most token insertions, deletions or substitutions, with trusted instructions immutable. For an explicit finite-state predicate and a randomized defense , certify , while preserving benign-task utility within of the undefended model. Determine the certificate's inference/sample cost.
Assumptions and scope
- Fix tokenizer, numeric/rounding semantics, parameter bit bounds, sampling rule and the known finite-state output predicate before certification. The predicate defines a restricted testable safety property; it is not an oracle for all harmful meanings. Inputs after editing have length at most n+k.
- Use a supplied benign prompt distribution and bounded task-loss function, with paid independent utility samples. Refusal is scored as a task failure when the task requires a substantive answer, preventing an always-refuse certificate from meeting the utility target.
- State white-box or query access, confidence 1-delta of randomized certification, and whether the guarantee is pointwise or distributional over base prompts. Count all model calls and preprocessing. Robust classification by a safety head does not by itself bound harmful generated responses.
What would settle it?
A sound nonvacuous defense/certificate with explicit n,k,H,|V| and confidence complexity on a stated useful model class, or matching computational/query lower bounds. Exhaustive finite enumeration supplies a baseline; attack-suite failure rates are not a certificate.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
Erase-and-check already certifies harmful-prompt detection relative to a base filter; SmoothLLM, CluCERT and output-oracle smoothing give further methods and conditional guarantees. The added target combines an explicit generated-output predicate, edit set and utility/cost constraint. Its architecture/defense design remains an editorial research framework. This is an editorial, source-motivated research target. The preliminary review through 2026-11-20 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Kumar et al. — Certifying LLM Safety against Adversarial Prompting ↗partial theory · source checked 2026-11-20
HTML Sections 3–7 and Theorem 4.1 checked: erase-and-check inherits harmful-prompt detection guarantees from a base filter; general infusion enumerates erased subsets. Benign utility is evaluated empirically and the guarantee is not unrestricted semantic output safety.
- Robey et al. — SmoothLLM: Defending Large Language Models Against Jailbreaking Attacks ↗related algorithm · source checked 2026-11-20
Primary abstract checked: perturb-and-aggregate defense with empirical robustness/utility trade-offs. Threat-model and instability assumptions matter; arbitrary semantic jailbreak immunity is not established.
- Wang et al. — CluCERT: Certifying LLM Robustness via Clustering-Guided Denoising Smoothing ↗partial theory · source checked 2026-11-20
Primary search abstract checked: semantic clustering and denoising seek tighter certificates and lower sampling cost. This is an existing efficiency baseline, not a certificate against arbitrary meanings or input transformations.
- Seferis et al. — Randomized Smoothing Meets Vision-Language Models ↗partial theory · source checked 2026-11-20
Primary search abstract checked: generated responses mapped to discrete classes by a bounded-error oracle, with sample/radius laws. Output-level smoothing already has theory conditional on the classifier assumptions.