How can AI agents trade and execute safely in financial markets? Our CEO Raul Heraud explored this question with fellow panelists at @wallet’s Beyond X in Singapore. Proud to support an evening bringing this conversation to the stage.

Trail of Bits published notes on 18 September on reviewing the Miden zkVM. The part worth reading is what happened before the review started. Security firms — the same firm included — have published a number of posts describing how they pointed an agent harness at a codebase. This post is about a stage earlier than that: "before code review even starts, agents now allow us to build custom tooling and formal models that improve the quality and depth of our reviews." Miden has its own assembly language, MASM, and very little tooling around it. The firm had six months of lead time, because the implementation was not yet feature complete. It spent them having its agents build an LSP server, a decompiler, a static analysis engine, and a Lean model of the VM executor. None of that was the review. Applied during the review that followed, those tools produced concrete results. The static analysis identified over 400 unique locations where type validation could be improved, all of them reachable from the library's public API, as well as one high-severity finding: a remainder supplied by the prover was never validated before being passed to a 32-bit subtraction — "a malicious prover could exploit this to forge Falcon signatures and drain any Miden account controlled by a Falcon key pair." The Lean modeling effort produced 95 machine-checked correctness proofs covering all the binary arithmetic in the core library, and surfaced two more bugs the existing unit tests had not caught. The detail worth sitting with is smaller than any of those. The decompiler was the single largest piece of the work, over 100 agent-generated commits across months. The firm's own account of where the benefit landed: "The main benefit of this work turned out to be the decompiler's internal analysis frameworks and intermediate representation, which we could reuse for static analysis, rather than the full decompilation pipeline." The post is direct about why work like this did not happen before. It is "highly exploratory in nature, and the end results and potential payoff may be difficult to predict," which makes it "hard to sell clients on them in advance." What has changed, on the post's own account, is the cost of a miss: "Today, a failed side project only costs tokens." The payoff is no easier to predict than it was. What a wrong guess costs is smaller. One thing about this particular case is worth noting on its own. The main benefit ended up somewhere the single largest piece of the work had not been aimed at. That is a payoff shape easier to recognize once the work is finished than to argue for before it starts. blog.trailofbits.com/2026/09… #DeepSafe #Web3Security