Eric McCarthy
|
|
||||||
Bio
My driving interests are in applying formal methods to AI safety, and applying AI to accelerate formal methods.
I believe we must verify the software stacks used by AI models as soon as possible, to prevent theft and misuse as well as to prevent model self-exfiltration. To do that, we need to accelerate formal methods with AI agent help.
On the former I work on language security for ACL2, on formalization of programming language syntax, semantics, analysis, and transformation, such as formal derivations of C code, and formal verification of zero-knowledge circuits such as R1CS and AIR.
On the latter I have been improving an MCP server for ACL2, and I worked on a government-funded project where we trained an early transformer model to give advice to ACL2 users on how to fix failed proofs.
Since different formal methods have different capabilities, I believe for critical projects it is important to apply a variety of methods to achieve higher assurance. My main theorem prover is ACL2, and I hope to work on connecting other provers to ACL2.
I am currently working on a project to verify Rust code. I developed an ACL2 backend for Aeneas, used it to extract AES-128 from RustCrypto, and proved the extracted code equivalent to a well-validated executable specification. I also developed an interpreter of a subset of MIR (Rust’s Mid-level IR) in ACL2, and tested it against the specification. This work is described in this paper accepted to the 2026 ACL2 Workshop. For more detail on the paper, see this explainer.
I am also working on a government-funded project where we are synthesizing network filters in C (with proofs of correctness) to protect devices that have unpatched vulnerabilities. I am also working on a government-funded project where we are formalizing the Remora → MLIR compiler. A Remora Tutorial.
My past work, inside and outside Kestrel, includes:
-
Development of the Syntheto formal language and safe communication between its IDE and an ACL2 server, on the DARPA MIDAS project.
-
Formalization of parts of Aleo.
-
Formalization of parts of Zcash that are also used by Tezos.
-
Formalization of parts of the Yul intermediate language for Ethereum.
-
The ACL2 Ethereum project.
-
The DerivationMiner project, where I developed (in Python) the “big code” pipeline we used to process and run machine learning on a corpus of 23 TB. I saved Java bytecode and artifacts to a Titan graph database, and ran NMF for dimension reduction. I also developed our code similarity search tool. I would like to apply these “big code” tools to corpora of smart contract code. I also learned how to use the ACL2 theorem prover and did some formal verification as part of this project.
-
The VIBRANCE project, where we automatically hardened Java bytecode against attacks from tainted inputs.
-
With Decidable Software, developed a tool using Software Refinery to automate a large source code port from a proprietary PL/I-like language to standard COBOL for a French customer. Recruited two additional full-time developers for this project.
-
Architect and lead developer for tools used in reinspection service, written in Python and Java. (Reasoning, post-Y2K era)
-
Implemented a distributed processing system for large analysis jobs. (Reasoning, Y2K era)
-
UX/UI design and implementation for Y2K analysts. (Reasoning, Y2K era)
-
Extended Intervista UI toolkit and Dialect grammar compiler to handle Japanese text. (Reasoning, Software Refinery era)
-
Software Refinery core developer for 15 years, bringing into production the Refine IDE, Dialect grammar compiler, Intervista UI toolkit, REFINE/Ada, REFINE/C, REFINE/COBOL, and REFINE/FORTRAN static analysis products. Programming in Lisp and Refine.
I received a degree in Mathematical Sciences from Stanford University in 1984.