BosqueLanguage (@BosqueLanguage) profile picture

Twitter account for the Bosque Development Platform. Tweets by Mark Marron.

Lexington, KY
Joined April 2019
Photo by BosqueLanguage (@BosqueLanguage)Photo by BosqueLanguage (@BosqueLanguage)Photo by BosqueLanguage (@BosqueLanguage)Photo by BosqueLanguage (@BosqueLanguage)
BosqueLanguage retweeted
Delighted to share the the work of my team at @GoogleDeepMind: formally verifying a real-world implementation of the #Signal cryptographic protocol at the machine-code level. The proof was done using AlphaProof Code, our new AI framework for formal specification and verification.
Our teams put Gemini 4 Carbon to work on a major security challenge: mathematically proving the code underlying the Signal protocol. Using AlphaProof Code, @GoogleDeepMind researchers guided agents to verify the protocol down to compiled device code, closing the gap where real bugs hide. The result: - Gemini 4 Carbon generated over 6M lines of machine-checked Lean proofs - years of expert manual verification compressed to just weeks ~1% of the codebase required human review (Lean automatically checked the rest) This is helping unlock scalable formal verification for building secure software, laying the groundwork to fundamentally shift the advantage to defenders. We’re open sourcing the results for the security community.
2
12
1
128
6,976
100% agree, as much as possible specs should in the same language as the code -- avoids a 2 language problem and makes easy runtime checking. Our recent paper -- just on specs might also be relevant (dl.acm.org/doi/abs/10.1145/3…) extending them to make modeling environmental behaviors easy/natural for humans, agents, and (of course) formal methods!
1
131
BosqueLanguage retweeted
Beam’s model technical report is now available. Read the full report to learn about the methodology our team used to train Beam end-to-end in-house, here: reflection.ai/documents/beam…
21
68
32
476
70,327
+1 and I like Marcel's paper too. We are also having a lot of success with the inverse direction for integration -- having the fuzzer drive the agent (arxiv.org/pdf/2510.19777)!
1
115
I like using agents to help with coding but I DO NOT like having them run riot through my codebase! Bosque is trying to use syntactic holes to direct where agents write code -- and it feels better. I also like to try out flows before all the code is complete and holes work for this too -- better than lots of assert(false, "not implemented) statements. A colleague of mine noted that you don't need to actually implement this code. Instead you can dynamically build a table of the right input/output values and use those to simulate the missing bits. Now you can imagine not "pre-writing" and then running an application but instead, having needed behaviors generated Just-In-Time, at runtime! In fact an AI agent can lazily fill in values, maybe write code, and if the results are not quite what you want, rewind time and try a different branch of what could happen. We wrote up a short abstract to help us think this through, and will be presenting it at VMIL @VMIL_Workshop (with @splashcon). If the idea of Agenticly-JITed sowtware development sounds check out the abstract and share your thoughts! arxiv.org/abs/2609.10248
1
2
46
Replying to @nateberkopec
Glad to hear that! Always happy to answer any questions or discuss in more detail. Will be talking about this at the SpecOps conference in Oakland next month as well: conf.researchr.org/home/spla…
3
128
👍hope it is useful, and always happy to answer questions/discuss more
1
31
Replying to @nateberkopec
Great to see this term used. I believe it is a foundational concept for building trustworthy systems. A big push recently is integrating it as the core of a specification language for agentic APIs. Built in from the start instead of added incrementally and ad-hoc. arxiv.org/pdf/2606.15084
2
26
4,217
You might also be interested in some test generation for APIs (or a PBT harness) that a student of mine has been leading. Outperforms coverage guided (and definitely randomized) fuzzers by a large margin in coverage and errors discovered (2x more detected over SotA). It combines combinatorial testing as mentioned in the post (k-combinations) with a much stronger way to sample (and generate) the value distributions using LLMs. arxiv.org/pdf/2510.19777
1
2
33
Replying to @ShriramKMurthi
No worries, it was definitely a fun project and I learned a bunch from talking developers and hearing about their pain-points. Happy to answer any questions or discuss some of the feedback we got from the Node.js community on the work. Personally, I am getting ready to take this info and build an async runtime with structured concurrency (plus priorities and causal dependency tracking) for Bosque. Cancellation and error prop. are definitely challenging issues so I was great to see them covered in this OOPSLA paper!
1
1
16
Replying to @ShriramKMurthi
Hey, very interesting. You might be interested in a paper I worked on a few years ago when working with the Chakra team at Microsoft on the semantics of async Node.js Depending on the async API you might get multiple different semantics as things like I/O, timeouts, and promises have different dispatch queues and priorities! Good engineering reasons but makes for a complex bit of implicit priorities, possible execution orders, and semantics. dl.acm.org/doi/abs/10.1145/3…
1
1
40
BosqueLanguage retweeted
AI oversight and control is the most difficult problem I can think of. Come help me try to solve it
1
8
1
12
2,919
BosqueLanguage retweeted
We're hiring! Come help us build tools for overseeing powerful AIs oath.tech/jobs/
3
5
2
17
3,541
Well Chain-of-Thought (CoT) has been, and perhaps the end of trying to monitor it, has been a topic of conversation today, and honestly, I am happy to say good riddance to this approach! Trying to read the inner monolouge (if that is even what CoT faithfully represents) and use that information to drive correct behavior is fundamentally a round-about and fragile approach to achieving what we want – a correct outcome. Rather than trying to accomplish correctness by shaping inner thoughts (or processes) into perfection, we much more robust methods for ensuring desired outcomes. We ask for proof of satisfaction in addition to simply producing an answer or artifact!
1
120
Basically, tried to get rid of all the common attack vectors that I saw working on JavaScript engines (lower level) and with investment banks (mainly data exfiltration) + future attacks based on supplychains, trust, and implicit (unvalidated) assumptions.
11
The easily part is more interesting. Bosque is built with a pit-of-succeess model. So, there are many simple ways to handle common security properties. One example is (ReDoS free) regex validated strings as a builtin feature. Making it easy to eliminate and validate SQL injection, path escape, and other common un-trusted input attacks. Another is the ability to specify resource use sandboxes using URIs. So, in the linked blog saying that a agent can only read files in the \/tmp/myapp/**\ directory. I am also a big fan of the design of the immutable (preventing an agent from forging something) event log that is strictly more expressive than LTL/CTL and easier to write checks against. And also lets us say useful things like -- any payment and agent makes is under $100 USD /\ has explicit user authorization.
1
13
Bosque takes a very simple approach to the first as the specification language is the programming language. So, any property that is computable can be specified (and discharged via a range of proof systems). We built prototype test generation and small model verification -- but translation to a proof-oriented language is a direct process.
1
2
13
I think there are 2 parts to this answer. First is theoretically what properties can be specified/proved and Second what can be **easily** specified/proved.
1
1
19
Hey, I am on the other side of that bet! You definitely can prove security properties with formal methods -- just historically hasn't been a focus there. But is definitely something that we have been investigating with the Bosque language (bosquelanguage.github.io/202…). See also the Cedar project at AWS (dl.acm.org/doi/10.1145/36498…). Interestingly focusing on verifying agent actions explicitly, instead of verifying security of a full OS, is actually easier in many ways. It is just checking a specific (and smaller) sequence of proposed actions satisfies the specified security policy. This is much more tractable for formal methods -- even today.
1
1
16