Every major technical invention resurfaces the dual-use dilemma—the new technology has the potential to be used for good as well as for harm. Generative AI (GenAI) techniques, such as large language models (LLMs) and diffusion models, have shown remarkable capabilities (e.g., in-context learning, code-completion, and text-to-image generation and editing). However, GenAI can be used just as well by attackers to generate new attacks and increase the velocity and efficacy of existing attacks. This monograph reports the findings of a workshop held at Google (co-organized by Stanford University and the University of Wisconsin-Madison) on the dual-use dilemma posed by GenAI. This work is not meant to be comprehensive, but is rather an attempt to synthesize some of the interesting findings from the workshop. We discuss short-term and long-term goals for the community on this topic. We hope this work provides both a launching point for a discussion on this important topic as well as interesting problems that the research community can work to address.

Common data types like dates, addresses, phone numbers and tables can have multiple textual representations, and many heavily-used languages, such as SQL, come in several dialects. These variations can cause data to be misinterpreted, leading to silent data corruption, failure of data processing systems, or even security vulnerabilities. Saggitarius is a new language and system designed to help programmers reason about the format of data, by describing grammatical domains—that is, sets of context-free grammars that describe the many possible representations of a datatype. We describe the design of Saggitarius via example and provide a relational semantics. We show how Saggitarius may be used to analyze a data set: given example data, it uses an algorithm based on semi-ring parsing and MaxSAT to infer which grammar in a given domain best matches that data. We evaluate the effectiveness of the algorithm on a benchmark suite of 110 example problems, and we demonstrate that our system typically returns a satisfying grammar within a few seconds with only a small number of examples. We also delve deeper into a more extensive case study on using Saggitarius for CSV dialect detection. Despite being general-purpose, we find that Saggitarius offers comparable results to hand-tuned, specialized tools; in the case of CSV, it infers grammars for 84% of benchmarks within 60 seconds, and has comparable accuracy to custom-built dialect detection tools.

A lens is a single program that specifies two data transformations at once: one transformation converts data from source format to target format and a second transformation inverts the process. Over the past decade, researchers have developed many different kinds of lenses with different properties. One class of such languages operate over regular languages. In other words, these lenses convert strings drawn from one regular language to strings drawn from another regular language (and back again). In this paper, we define a more powerful language of lenses, which we call match-reference lenses, that is capable of translating between non-regular formats that contain repeated substrings, which is a primitive form of dependency. To define the non-regular formats themselves, we develop a new language, match-reference regular expressions, which are regular expressions that can bind variables to substrings and use those substrings repeatedly. These match-reference regular expressions are closely related to the familiar “back-references“ that can be found in traditional regular expression packages, but are redesigned to adhere to conventional programming language lexical scoping conventions and to interact smoothly with lens language infrastructure. We define the semantics of match-reference regular expressions and match-reference lenses. We also define a new kind of automaton, the match-reference regex automaton system (MRRAS), for deciding string membership in the language match-reference regular expressions. We illustrate our definitions with a variety of examples.

Parsers are security-critical components of many software systems, and verified parsing therefore has a key role to play in secure software design. However, existing verified parsers for context-free grammars are limited in their expressiveness, termination properties, or performance characteristics. They are only compatible with a restricted class of grammars, they are not guaranteed to terminate on all inputs, or they are not designed to be performant on grammars for real-world programming languages and data formats. In this work, we present CoStar, a verified parser that addresses these limitations. The parser is implemented with the Coq Proof Assistant and is based on the ALL(*) parsing algorithm. CoStar is sound and complete for all non-left-recursive grammars; it produces a correct parse tree for its input whenever such a tree exists, and it correctly detects ambiguous inputs. CoStar also provides strong termination guarantees; it terminates without error on all inputs when applied to a non-left-recursive grammar. Finally, CoStar achieves linear-time performance on a range of unambiguous grammars for commonly used languages and data formats.

Lexers and parsers are often used as front ends to connect input from the outside world with the internals of a larger software system. These front ends are natural targets for attackers who wish to compromise the larger system. A formally verified tool that performs mechanized lexical analysis would render attacks on these front ends less effective. In this paper we present Verbatim, an executable lexer that is implemented and verified with the Coq Proof Assistant. We prove that Verbatim is correct with respect to a standard lexer specification. We also analyze its theoretical complexity and give results of an empirical performance evaluation. All correctness proofs have been mechanized in Coq.

Lenses are programs that can be run both “front to back” and “back to front,” allowing updates to either their source or their target data to be transferred in both directions. Since their introduction by Foster et al., lenses have been extensively studied, extended, and applied. Recent work has also demonstrated how techniques from type-directed program synthesis can be used to efficiently synthesize a simple class of lenses—so-called bijective lenses over string data—given a pair of types (regular expressions) and a small number of examples. We extend this synthesis algorithm to a much broader class of lenses, called simple symmetric lenses, including all bijective lenses, all of the popular category of “asymmetric” lenses, and a rich subset of the more powerful “symmetric lenses” proposed by Hofmann et al. Intuitively, simple symmetric lenses allow some information to be present on one side but not the other and vice versa. They are of independent theoretical interest, being the largest class of symmetric lenses that do not rely on persistent internal state. Synthesizing simple symmetric lenses is substantially more challenging than synthesizing bijective lenses: Since some of the information on each side can be “disconnected” from the other side, there will, in general, be many lenses that agree with a given example. To guide the search process, we use stochastic regular expressions and ideas from information theory to estimate the amount of information propagated by a candidate lens, generally preferring lenses that propagate more information, as well as user annotations marking parts of the source and target data structures as either irrelevant or essential. We describe an implementation of simple symmetric lenses and our synthesis procedure as extensions to the Boomerang language. We evaluate its performance on 48 benchmark examples drawn from Flash Fill, Augeas, the bidirectional programming literature, and electronic file format synchronization tasks. Our implementation can synthesize each of these lenses in under 30 seconds.

An LL(1) parser is a recursive descent algorithm that uses a single token of lookahead to build a grammatical derivation for an input sequence. We present an LL(1) parser generator that, when applied to grammar G, produces an LL(1) parser for G if such a parser exists. We use the Coq Proof Assistant to verify that the generator and the parsers that it produces are sound and complete, and that they terminate on all inputs without using fuel parameters. As a case study, we extract the tool’s source code and use it to generate a JSON parser. The generated parser runs in linear time; it is two to four times slower than an unverified parser for the same grammar.

Bidirectional transformations between different data representations occur frequently in modern software systems. They appear as serializers and deserializers, as parsers and pretty printers, as database views and view updaters, and as a multitude of different kinds of ad hoc data converters. Manually building bidirectional transformations—by writing two separate functions that are intended to be inverses—is tedious and error prone. A better approach is to use a domain-specific language in which both directions can be written as a single expression. However, these domain-specific languages can be difficult to program in, requiring programmers to manage fiddly details while working in a complex type system. We present an alternative approach. Instead of coding transformations manually, we synthesize them from declarative format descriptions and examples. Specifically, we present Optician, a tool for type-directed synthesis of bijective string transformers. The inputs to Optician are a pair of ordinary regular expressions representing two data formats and a few concrete examples for disambiguation. The output is a well-typed program in Boomerang (a bidirectional language based on the theory of lenses). The main technical challenge involves navigating the vast program search space efficiently. In particular, and unlike most prior work on type-directed synthesis, our system operates in the context of a language with a rich equivalence relation on types (the theory of regular expressions). Consequently, program synthesis requires search in two dimensions: First, our synthesis algorithm must find a pair of “syntactically compatible types,” and second, using the structure of those types, it must find a type- and example-compliant term. Our key insight is that it is possible to reduce the size of this search space without losing any computational power by defining a new language of lenses designed specifically for synthesis. The new language is free from arbitrary function composition and operates only over types and terms in a new disjunctive normal form. We prove (1) our new language is just as powerful as a more natural, compositional, and declarative language and (2) our synthesis algorithm is sound and complete with respect to the new language. We also demonstrate empirically that our new language changes the synthesis problem from one that admits intractable solutions to one that admits highly efficient solutions, able to synthesize intricate lenses between complex file formats in seconds. We evaluate Optician on a benchmark suite of 39 examples that includes both microbenchmarks and realistic examples derived from other data management systems including Flash Fill, a tool for synthesizing string transformations in spreadsheets, and Augeas, a tool for bidirectional processing of Linux system configuration files.

For decades, formal methods have offered the promise of verified software that does not have exploitable bugs. Until recently, however, it has not been possible to verify software of sufficient complexity to be useful. Recently, that situation has changed. SeL4 is an open-source operating system microkernel efficient enough to be used in a wide range of practical applications. Its designers proved it to be fully functionally correct, ensuring the absence of buffer overflows, null pointer exceptions, use-after-free errors, etc., and guaranteeing integrity and confidentiality. The CompCert Verifying C Compiler maps source C programs to provably equivalent assembly language, ensuring the absence of exploitable bugs in the compiler. A number of factors have enabled this revolution, including faster processors, increased automation, more extensive infrastructure, specialized logics and the decision to co-develop code and correctness proofs rather than verify existing artefacts. In this paper, we explore the promise and limitations of current formal-methods techniques. We discuss these issues in the context of DARPA’s HACMS program, which had as its goal the creation of high-assurance software for vehicles, including quadcopters, helicopters and automobiles. This article is part of the themed issue ‘Verified trustworthy software systems’.