Zac Hatfield-Dodds — Page 2

arXiv2021

Deriving Semantics-Aware Fuzzers from Web API Schemas

Fuzzing -- whether generating or mutating inputs -- has found many bugs and security vulnerabilities in a wide range of domains. Stateful and highly structured web APIs present significant challenges to traditional fuzzing techniques, as execution feedback is usually limited to a response code instead of code coverage and vulnerabilities of interest include silent information-disclosure in addition to explicit errors. Our tool, Schemathesis, derives structure- and semantics-aware fuzzers from web API schemas in the OpenAPI or GraphQL formats, using property-based testing tools. Derived fuzzers can be incorporated into unit-test suites or run directly, with or without end-user customisation of data generation and semantic checks. We construct the most comprehensive evaluation of web API fuzzers to date, running eight fuzzers against sixteen real-world open source web services. OpenAPI schemas found in the wild have a long tail of rare features and complex structures. Of the tools we evaluated, Schemathesis was the only one to handle more than two-thirds of our target services without a fatal internal error. Schemathesis finds 1.4 times to 4.5 times more unique defects than the respectively second-best fuzzer for each target, and is the only fuzzer to find defects in four targets.

cs.CR↗

arXiv2021

A General Language Assistant as a Laboratory for Alignment

Given the broad capabilities of large language models, it should be possible to work towards a general-purpose, text-based assistant that is aligned with human values, meaning that it is helpful, honest, and harmless. As an initial foray in this direction we study simple baseline techniques and evaluations, such as prompting. We find that the benefits from modest interventions increase with model size, generalize to a variety of alignment evaluations, and do not compromise the performance of large models. Next we investigate scaling trends for several training objectives relevant to alignment, comparing imitation learning, binary discrimination, and ranked preference modeling. We find that ranked preference modeling performs much better than imitation learning, and often scales more favorably with model size. In contrast, binary discrimination typically performs and scales very similarly to imitation learning. Finally we study a `preference model pre-training' stage of training, with the goal of improving sample efficiency when finetuning on human preferences.

cs.CL↗